Blog

Agent memory remains a problem that enterprises want to fix, as agents forget some instructions or conversations the longer they run. 

Anthropic believes it has solved this issue for its Claude Agent SDK, developing a two-fold solution that allows an agent to work across different context windows.

“The core challenge of long-running agents is that they must work in discrete sessions, and each new session begins with no memory of what came before,” Anthropic wrote in a blog post. “Because context windows are limited, and because most complex projects cannot be completed within a single window, agents need a way to bridge the gap between coding sessions.”

Anthropic engineers proposed a two-fold approach for its Agent SDK: An initializer agent to set up the environment, and a coding agent to make incremental progress in each session and leave artifacts for the next.  

The agent memory problem

Since agents are built on foundation models, they remain constrained by the limited, although continually growing, context windows. For long-running agents, this could create a larger problem, leading the agent to forget instructions and behave abnormally while performing a task. Enhancing agent memory becomes essential for consistent, business-safe performance. 

Several methods emerged over the past year, all attempting to bridge the gap between context windows and agent memory. LangChain’s LangMem SDK, Memobase and OpenAI’s Swarm are examples of companies offering memory solutions. Research on agentic memory has also exploded recently, with proposed frameworks like Memp and the Nested Learning Paradigm from Google offering new alternatives to enhance memory. 

Many of the current memory frameworks are open source and can ideally adapt to different large language models (LLMs) powering agents. Anthropic’s approach improves its Claude Agent SDK. 

How it works

Anthropic identified that even though the Claude Agent SDK had context management capabilities and “should be possible for an agent to continue to do useful work for an arbitrarily long time,” it was not sufficient. The company said in its blog post that a model like Opus 4.5 running the Claude Agent SDK can “fall short of building a production-quality web app if it’s only given a high-level prompt, such as 'build a clone of claude.ai.'” 

The failures manifested in two patterns, Anthropic said. First, the agent tried to do too much, causing the model to run out of context in the middle. The agent then has to guess what happened and cannot pass clear instructions to the next agent. The second failure occurs later on, after some features have already been built. The agent sees progress has been made and just declares the job done. 

Anthropic researchers broke down the solution: Setting up an initial environment to lay the foundation for features and prompting each agent to make incremental progress towards a goal, while still leaving a clean slate at the end. 

This is where the two-part solution of Anthropic's agent comes in. The initializer agent sets up the environment, logging what agents have done and which files have been added. The coding agent will then ask models to make incremental progress and leave structured updates. 

“Inspiration for these practices came from knowing what effective software engineers do every day,” Anthropic said. 

The researchers said they added testing tools to the coding agent, improving its ability to identify and fix bugs that weren’t obvious from the code alone. 

Future research

Anthropic noted that its approach is “one possible set of solutions in a long-running agent harness.” However, this is just the beginning stage of what could become a wider research area for many in the AI space. 

The company said its experiments to boost long-term memory for agents haven’t shown whether a single general-purpose coding agent works best across contexts or a multi-agent structure. 

Its demo also focused on full-stack web app development, so other experiments should focus on generalizing the results across different tasks.

“It’s likely that some or all of these lessons can be applied to the types of long-running agentic tasks required in, for example, scientific research or financial modeling,” Anthropic said. 

Hello, dear readers. Happy belated Thanksgiving and Black Friday!

This year has felt like living inside a permanent DevDay. Every week, some lab drops a new model, a new agent framework, or a new “this changes everything” demo. It’s overwhelming. But it’s also the first year I’ve felt like AI is finally diversifying — not just one or two frontier models in the cloud, but a whole ecosystem: open and closed, giant and tiny, Western and Chinese, cloud and local.

So for this Thanksgiving edition, here’s what I’m genuinely thankful for in AI in 2025 — the releases that feel like they’ll matter in 12–24 months, not just during this week’s hype cycle.

1. OpenAI kept shipping strong: GPT-5, GPT-5.1, Atlas, Sora 2 and open weights

As the company that undeniably birthed the "generative AI" era with its viral hit product ChatGPT in late 2022, OpenAI arguably had among the hardest tasks of any AI company in 2025: continue its growth trajectory even as well-funded competitors like Google with its Gemini models and other startups like Anthropic fielded their own highly competitive offerings.

Thankfully, OpenAI rose to the challenge and then some. Its headline act was GPT-5, unveiled in August as the next frontier reasoning model, followed in November by GPT-5.1 with new Instant and Thinking variants that dynamically adjust how much “thinking time” they spend per task.

In practice, GPT-5’s launch was bumpy — VentureBeat documented early math and coding failures and a cooler-than-expected community reaction in “OpenAI’s GPT-5 rollout is not going smoothly," but it quickly course corrected based on user feedback and, as a daily user of this model, I'm personally pleased with it and impressed with it.

At the same time, enterprises actually using the models are reporting solid gains. ZenDesk Global, for example, says GPT-5-powered agents now resolve more than half of customer tickets, with some customers seeing 80–90% resolution rates. That’s the quiet story: these models may not always impress the chattering classes on X, but they’re starting to move real KPIs.

On the tooling side, OpenAI finally gave developers a serious AI engineer with GPT-5.1-Codex-Max, a new coding model that can run long, agentic workflows and is already the default in OpenAI’s Codex environment. VentureBeat covered it in detail in “OpenAI debuts GPT-5.1-Codex-Max coding model and it already completed a 24-hour task internally.”

Then there’s ChatGPT Atlas, a full browser with ChatGPT baked into the chrome itself — sidebar summaries, on-page analysis, and search tightly integrated into regular browsing. It’s the clearest sign yet that “assistant” and “browser” are on a collision course.

On the media side, Sora 2 turned the original Sora video demo into a full video-and-audio model with better physics, synchronized sound and dialogue, and more control over style and shot structure, plus a dedicated Sora app with a full fledged social networking component, allowing any user to create their own TV network in their pocket.

Finally — and maybe most symbolically — OpenAI released gpt-oss-120B and gpt-oss-20B, open-weight MoE reasoning models under an Apache 2.0–style license. Whatever you think of their quality (and early open-source users have been loud about their complaints), this is the first time since GPT-2 that OpenAI has put serious weights into the public commons.

2. China’s open-source wave goes mainstream

If 2023–24 was about Llama and Mistral, 2025 belongs to China’s open-weight ecosystem.

A study from MIT and Hugging Face found that China now slightly leads the U.S. in global open-model downloads, largely thanks to DeepSeek and Alibaba’s Qwen family.

Highlights:

• DeepSeek-R1 dropped in January as an open-source reasoning model rivaling OpenAI’s o1, with MIT-licensed weights and a family of distilled smaller models. VentureBeat has followed the story from its release to its cybersecurity impact to performance-tuned R1 variants.

• Kimi K2 Thinking from Moonshot, a “thinking” open-source model that reasons step-by-step with tools, very much in the o1/R1 mold, and is positioned as the best open reasoning model so far in the world.

• Z.ai shipped GLM-4.5 and GLM-4.5-Air as “agentic” models, open-sourcing base and hybrid reasoning variants on GitHub.

• Baidu’s ERNIE 4.5 family arrived as a fully open-sourced, multimodal MoE suite under Apache 2.0, including a 0.3B dense model and visual “Thinking” variants focused on charts, STEM, and tool use.

• Alibaba’s Qwen3 line — including Qwen3-Coder, large reasoning models, and the Qwen3-VL series released over the summer and fall months of 2025 — continues to set a high bar for open weights in coding, translation, and multimodal reasoning, leading me to declare this past summer as "

  Qwen's summer."

VentureBeat has been tracking these shifts, including Chinese math and reasoning models like Light-R1-32B and Weibo’s tiny VibeThinker-1.5B, which beat DeepSeek baselines on shoestring training budgets.

If you care about open ecosystems or on-premise options, this is the year China’s open-weight scene stopped being a curiosity and became a serious alternative.

3. Small and local models grow up

Another thing I’m thankful for: we’re finally getting good small models, not just toys.

Liquid AI spent 2025 pushing its Liquid Foundation Models (LFM2) and LFM2-VL vision-language variants, designed from day one for low-latency, device-aware deployments — edge boxes, robots, and constrained servers, not just giant clusters. The newer LFM2-VL-3B targets embedded robotics and industrial autonomy, with demos planned at ROSCon.

On the big-tech side, Google’s Gemma 3 line made a strong case that “tiny” can still be capable. Gemma 3 spans from 270M parameters up through 27B, all with open weights and multimodal support in the larger variants.

The standout is Gemma 3 270M, a compact model purpose-built for fine-tuning and structured text tasks — think custom formatters, routers, and watchdogs — covered both in Google’s developer blog and community discussions in local-LLM circles.

These models may never trend on X, but they’re exactly what you need for privacy-sensitive workloads, offline workflows, thin-client devices, and “agent swarms” where you don’t want every tool call hitting a giant frontier LLM.

4. Meta + Midjourney: aesthetics as a service

One of the stranger twists this year: Meta partnered with Midjourney instead of simply trying to beat it.

In August, Meta announced a deal to license Midjourney’s “aesthetic technology” — its image and video generation stack — and integrate it into Meta’s future models and products, from Facebook and Instagram feeds to Meta AI features.

VentureBeat covered the partnership in “Meta is partnering with Midjourney and will license its technology for future models and products,” raising the obvious question: does this slow or reshape Midjourney’s own API roadmap? Still awaiting an answer there, but unfortunately, stated plans for an API release have yet to materialize, suggesting that it has.

For creators and brands, though, the immediate implication is simple: Midjourney-grade visuals start to show up in mainstream social tools instead of being locked away in a Discord bot. That could normalize higher-quality AI art for a much wider audience — and force rivals like OpenAI, Google, and Black Forest Labs to keep raising the bar.

5. Google’s Gemini 3 and Nano Banana Pro

Google tried to answer GPT-5 with Gemini 3, billed as its most capable model yet, with better reasoning, coding, and multimodal understanding, plus a new Deep Think mode for slow, hard problems.

VentureBeat’s coverage, “Google unveils Gemini 3 claiming the lead in math, science, multimodal and agentic AI,” framed it as a direct shot at frontier benchmarks and agentic workflows.

But the surprise hit is Nano Banana Pro (Gemini 3 Pro Image), Google’s new flagship image generator. It specializes in infographics, diagrams, multi-subject scenes, and multilingual text that actually renders legibly across 2K and 4K resolutions.

In the world of enterprise AI — where charts, product schematics, and “explain this system visually” images matter more than fantasy dragons — that’s a big deal.

6. Wild cards I’m keeping an eye on

A few more releases I’m thankful for, even if they don’t fit neatly into one bucket:

• Black Forest Labs’ Flux.2 image models, which launched just earlier this week with ambitions to challenge both Nano Banana Pro and Midjourney on quality and control. VentureBeat dug into the details in “Black Forest Labs launches Flux.2 AI image models to challenge Nano Banana Pro and Midjourney."

• Anthropic’s Claude Opus 4.5, a new flagship that aims for cheaper, more capable coding and long-horizon task execution, covered in “Anthropic’s Claude Opus 4.5 is here: Cheaper AI, infinite chats, and coding skills that beat humans."

• A steady drumbeat of open math/reasoning models — from Light-R1 to VibeThinker and others — that show you don’t need $100M training runs to move the needle.

Last thought (for now)

If 2024 was the year of “one big model in the cloud,” 2025 is the year the map exploded: multiple frontiers at the top, China taking the lead in open models, small and efficient systems maturing fast, and creative ecosystems like Midjourney getting pulled into big-tech stacks.

I’m thankful not just for any single model, but for the fact that we now have options — closed and open, local and hosted, reasoning-first and media-first. For journalists, builders, and enterprises, that diversity is the real story of 2025.

Happy holidays and best to you and your loved ones!

Researchers at Alibaba’s Tongyi Lab have developed a new framework for self-evolving agents that create their own training data by exploring their application environments. The framework, AgentEvolver, uses the knowledge and reasoning capabilities of large language models for autonomous learning, addressing the high costs and manual effort typically required to gather task-specific datasets.

Experiments show that compared to traditional reinforcement learning–based frameworks, AgentEvolver is more efficient at exploring its environment, makes better use of data, and adapts faster to application environments. For the enterprise, this is significant because it lowers the barrier to training agents for bespoke applications, making powerful, custom AI assistants more accessible to a wider range of organizations.

The high cost of training AI agents

Reinforcement learning has become a major paradigm for training LLMs to act as agents that can interact with digital environments and learn from feedback. However, developing agents with RL faces fundamental challenges. First, gathering the necessary training datasets is often prohibitively expensive, requiring significant manual labor to create examples of tasks, especially in novel or proprietary software environments where there are no available off-the-shelf datasets.

Second, the RL techniques commonly used for LLMs require the model to run through a massive number of trial-and-error attempts to learn effectively. This process is computationally costly and inefficient. As a result, training capable LLM agents through RL remains laborious and expensive, limiting their deployment in custom enterprise settings.

How AgentEvolver works

The main idea behind AgentEvolver is to give models greater autonomy in their own learning process. The researchers describe it as a “self-evolving agent system” designed to “achieve autonomous and efficient capability evolution through environmental interaction.” It uses the reasoning power of an LLM to create a self-training loop, allowing the agent to continuously improve by directly interacting with its target environment without needing predefined tasks or reward functions.

“We envision an agent system where the LLM actively guides exploration, task generation, and performance refinement,” the researchers wrote in their paper.

The self-evolution process is driven by three core mechanisms that work together.

The first is self-questioning, where the agent explores its environment to discover the boundaries of its functions and identify useful states. It’s like a new user clicking around an application to see what’s possible. Based on this exploration, the agent generates its own diverse set of tasks that align with a user’s general preferences. This reduces the need for handcrafted datasets and allows the agent and its tasks to co-evolve, progressively enabling it to handle more complex challenges. 

According to Yunpeng Zhai, researcher at Alibaba and co-author of the paper, who spoke to VentureBeat, the self-questioning mechanism effectively turns the model from a “data consumer into a data producer,” dramatically reducing the time and cost required to deploy an agent in a proprietary environment.

The second mechanism is self-navigating, which improves exploration efficiency by reusing and generalizing from past experiences. AgentEvolver extracts insights from both successful and unsuccessful attempts and uses them to guide future actions. For example, if an agent tries to use an API function that doesn't exist in an application, it registers this as an experience and learns to verify the existence of functions before attempting to use them in the future.

The third mechanism, self-attributing, enhances learning efficiency by providing more detailed feedback. Instead of just a final success or failure signal (a common practice in RL that can result in sparse rewards), this mechanism uses an LLM to assess the contribution of each individual action in a multi-step task. It retrospectively determines whether each step contributed positively or negatively to the final outcome, giving the agent fine-grained feedback that accelerates learning. 

This is crucial for regulated industries where how an agent solves a problem is as important as the result. “Instead of rewarding a student only for the final answer, we also evaluate the clarity and correctness of each step in their reasoning,” Zhai explained. This improves transparency and encourages the agent to adopt more robust and auditable problem-solving patterns.

“By shifting the training initiative from human-engineered pipelines to LLM-guided self-improvement, AgentEvolver establishes a new paradigm that paves the way toward scalable, cost-effective, and continually improving intelligent systems,” the researchers state.

The team has also developed a practical, end-to-end training framework that integrates these three mechanisms. A key part of this foundation is the Context Manager, a component that controls the agent's memory and interaction history. While today's benchmarks test a limited number of tools, real enterprise environments can involve thousands of APIs. 

Zhai acknowledges this is a core challenge for the field, but notes that AgentEvolver was designed to be extended. “Retrieval over extremely large action spaces will always introduce computational challenges, but AgentEvolver’s architecture provides a clear path toward scalable tool reasoning in enterprise settings,” he said.

A more efficient path to agent training

To measure the effectiveness of their framework, the researchers tested it on AppWorld and BFCL v3, two benchmarks that require agents to perform long, multi-step tasks using external tools. They used models from Alibaba’s Qwen2.5 family (7B and 14B parameters) and compared their performance against a baseline model trained with GRPO, a popular RL technique used to develop reasoning models like DeepSeek-R1.

The results showed that integrating all three mechanisms in AgentEvolver led to substantial performance gains. For the 7B model, the average score improved by 29.4%, and for the 14B model, it increased by 27.8% over the baseline. The framework consistently enhanced the models' reasoning and task-execution capabilities across both benchmarks. The most significant improvement came from the self-questioning module, which autonomously generates diverse training tasks and directly addresses the data scarcity problem.

The experiments also demonstrated that AgentEvolver can efficiently synthesize a large volume of high-quality training data. The tasks generated by the self-questioning module proved diverse enough to achieve good training efficiency even with a small amount of data.

For enterprises, this provides a path to creating agents for bespoke applications and internal workflows while minimizing the need for manual data annotation. By providing high-level goals and letting the agent generate its own training experiences, organizations can develop custom AI assistants more simply and cost-effectively.

“This combination of algorithmic design and engineering pragmatics positions AgentEvolver as both a research vehicle and a reusable foundation for building adaptive, tool-augmented agents,” the researchers conclude.

Looking ahead, the ultimate goal is much bigger. “A truly ‘singular model’ that can drop into any software environment and master it overnight is certainly the holy grail of agentic AI,” Zhai said. “We see AgentEvolver as a necessary step in that direction.” While that future still requires breakthroughs in model reasoning and infrastructure, self-evolving approaches are paving the way.

This weekend, Andrej Karpathy, the former director of AI at Tesla and a founding member of OpenAI, decided he wanted to read a book. But he did not want to read it alone. He wanted to read it accompanied by a committee of artificial intelligences, each offering its own perspective, critiquing the others, and eventually synthesizing a final answer under the guidance of a "Chairman."

To make this happen, Karpathy wrote what he called a "vibe code project" — a piece of software written quickly, largely by AI assistants, intended for fun rather than function. He posted the result, a repository called "LLM Council," to GitHub with a stark disclaimer: "I’m not going to support it in any way… Code is ephemeral now and libraries are over."

Yet, for technical decision-makers across the enterprise landscape, looking past the casual disclaimer reveals something far more significant than a weekend toy. In a few hundred lines of Python and JavaScript, Karpathy has sketched a reference architecture for the most critical, undefined layer of the modern software stack: the orchestration middleware sitting between corporate applications and the volatile market of AI models.

As companies finalize their platform investments for 2026, LLM Council offers a stripped-down look at the "build vs. buy" reality of AI infrastructure. It demonstrates that while the logic of routing and aggregating AI models is surprisingly simple, the operational wrapper required to make it enterprise-ready is where the true complexity lies.

How the LLM Council works: Four AI models debate, critique, and synthesize answers

To the casual observer, the LLM Council web application looks almost identical to ChatGPT. A user types a query into a chat box. But behind the scenes, the application triggers a sophisticated, three-stage workflow that mirrors how human decision-making bodies operate.

First, the system dispatches the user’s query to a panel of frontier models. In Karpathy’s default configuration, this includes OpenAI’s GPT-5.1, Google’s Gemini 3.0 Pro, Anthropic’s Claude Sonnet 4.5, and xAI’s Grok 4. These models generate their initial responses in parallel.

In the second stage, the software performs a peer review. Each model is fed the anonymized responses of its counterparts and asked to evaluate them based on accuracy and insight. This step transforms the AI from a generator into a critic, forcing a layer of quality control that is rare in standard chatbot interactions.

Finally, a designated "Chairman LLM" — currently configured as Google’s Gemini 3 — receives the original query, the individual responses, and the peer rankings. It synthesizes this mass of context into a single, authoritative answer for the user.

Karpathy noted that the results were often surprising. "Quite often, the models are surprisingly willing to select another LLM's response as superior to their own," he wrote on X (formerly Twitter). He described using the tool to read book chapters, observing that the models consistently praised GPT-5.1 as the most insightful while rating Claude the lowest. However, Karpathy’s own qualitative assessment diverged from his digital council; he found GPT-5.1 "too wordy" and preferred the "condensed and processed" output of Gemini.

FastAPI, OpenRouter, and the case for treating frontier models as swappable components

For CTOs and platform architects, the value of LLM Council lies not in its literary criticism, but in its construction. The repository serves as a primary document showing exactly what a modern, minimal AI stack looks like in late 2025.

The application is built on a "thin" architecture. The backend uses FastAPI, a modern Python framework, while the frontend is a standard React application built with Vite. Data storage is handled not by a complex database, but by simple JSON files written to the local disk.

The linchpin of the entire operation is OpenRouter, an API aggregator that normalizes the differences between various model providers. By routing requests through this single broker, Karpathy avoided writing separate integration code for OpenAI, Google, and Anthropic. The application does not know or care which company provides the intelligence; it simply sends a prompt and awaits a response.

This design choice highlights a growing trend in enterprise architecture: the commoditization of the model layer. By treating frontier models as interchangeable components that can be swapped by editing a single line in a configuration file — specifically the COUNCIL_MODELS list in the backend code — the architecture protects the application from vendor lock-in. If a new model from Meta or Mistral tops the leaderboards next week, it can be added to the council in seconds.

What's missing from prototype to production: Authentication, PII redaction, and compliance

While the core logic of LLM Council is elegant, it also serves as a stark illustration of the gap between a "weekend hack" and a production system. For an enterprise platform team, cloning Karpathy’s repository is merely step one of a marathon.

A technical audit of the code reveals the missing "boring" infrastructure that commercial vendors sell for premium prices. The system lacks authentication; anyone with access to the web interface can query the models. There is no concept of user roles, meaning a junior developer has the same access rights as the CIO.

Furthermore, the governance layer is nonexistent. In a corporate environment, sending data to four different external AI providers simultaneously triggers immediate compliance concerns. There is no mechanism here to redact Personally Identifiable Information (PII) before it leaves the local network, nor is there an audit log to track who asked what.

Reliability is another open question. The system assumes the OpenRouter API is always up and that the models will respond in a timely fashion. It lacks the circuit breakers, fallback strategies, and retry logic that keep business-critical applications running when a provider suffers an outage.

These absences are not flaws in Karpathy’s code — he explicitly stated he does not intend to support or improve the project — but they define the value proposition for the commercial AI infrastructure market.

Companies like LangChain, AWS Bedrock, and various AI gateway startups are essentially selling the "hardening" around the core logic that Karpathy demonstrated. They provide the security, observability, and compliance wrappers that turn a raw orchestration script into a viable enterprise platform.

Why Karpathy believes code is now "ephemeral" and traditional software libraries are obsolete

Perhaps the most provocative aspect of the project is the philosophy under which it was built. Karpathy described the development process as "99% vibe-coded," implying he relied heavily on AI assistants to generate the code rather than writing it line-by-line himself.

"Code is ephemeral now and libraries are over, ask your LLM to change it in whatever way you like," he wrote in the repository’s documentation.

This statement marks a radical shift in software engineering capability. Traditionally, companies build internal libraries and abstractions to manage complexity, maintaining them for years. Karpathy is suggesting a future where code is treated as "promptable scaffolding" — disposable, easily rewritten by AI, and not meant to last.

For enterprise decision-makers, this poses a difficult strategic question. If internal tools can be "vibe coded" in a weekend, does it make sense to buy expensive, rigid software suites for internal workflows? Or should platform teams empower their engineers to generate custom, disposable tools that fit their exact needs for a fraction of the cost?

When AI models judge AI: The dangerous gap between machine preferences and human needs

Beyond the architecture, the LLM Council project inadvertently shines a light on a specific risk in automated AI deployment: the divergence between human and machine judgment.

Karpathy’s observation that his models preferred GPT-5.1, while he preferred Gemini, suggests that AI models may have shared biases. They might favor verbosity, specific formatting, or rhetorical confidence that does not necessarily align with human business needs for brevity and accuracy.

As enterprises increasingly rely on "LLM-as-a-Judge" systems to evaluate the quality of their customer-facing bots, this discrepancy matters. If the automated evaluator consistently rewards "wordy and sprawled" answers while human customers want concise solutions, the metrics will show success while customer satisfaction plummets. Karpathy’s experiment suggests that relying solely on AI to grade AI is a strategy fraught with hidden alignment issues.

What enterprise platform teams can learn from a weekend hack before building their 2026 stack

Ultimately, LLM Council acts as a Rorschach test for the AI industry. For the hobbyist, it is a fun way to read books. For the vendor, it is a threat, proving that the core functionality of their products can be replicated in a few hundred lines of code.

But for the enterprise technology leader, it is a reference architecture. It demystifies the orchestration layer, showing that the technical challenge is not in routing the prompts, but in governing the data.

As platform teams head into 2026, many will likely find themselves staring at Karpathy’s code, not to deploy it, but to understand it. It proves that a multi-model strategy is not technically out of reach. The question remains whether companies will build the governance layer themselves or pay someone else to wrap the "vibe code" in enterprise-grade armor.

It's not just Google's Gemini 3, Nano Banana Pro, and Anthropic's Claude Opus 4.5 we have to be thankful for this year around the Thanksgiving holiday here in the U.S.

No, today the German AI startup Black Forest Labs released FLUX.2, a new image generation and editing system complete with four different models designed to support production-grade creative workflows.

FLUX.2 introduces multi-reference conditioning, higher-fidelity outputs, and improved text rendering, and it expands the company’s open-core ecosystem with both commercial endpoints and open-weight checkpoints.

While Black Forest Labs previously launched with and made a name for itself on open source text-to-image models in its Flux family, today's release includes one fully open-source component: the Flux.2 VAE, available now under the Apache 2.0 license.

Four other models of varying size and uses — Flux.2 [Pro], Flux.2 [Flex], and Flux.2 [Dev] —are not open source; Pro and Flex remain proprietary hosted offerings, while Dev is an open-weight downloadable model that requires a commercial license obtained directly from Black Forest Labs for any commercial use. An upcoming open-source model is Flux.2 [Klein], which will also be released under Apache 2.0 when available.

But the open source Flux.2 VAE, or variational autoencoder, is important and useful to enterprises for several reasons. This is a module that compresses images into a latent space and reconstructs them back into high-resolution outputs; in Flux.2, it defines the latent representation used across the multiple (four total, see blow) model variants, enabling higher-quality reconstructions, more efficient training, and 4-megapixel editing.

Because this VAE is open and freely usable, enterprises can adopt the same latent space used by BFL’s commercial models in their own self-hosted pipelines, gaining interoperability between internal systems and external providers while avoiding vendor lock-in.

The availability of a fully open, standardized latent space also enables practical benefits beyond media-focused organizations. Enterprises can use an open-source VAE as a stable, shared foundation for multiple image-generation models, allowing them to switch or mix generators without reworking downstream tools or workflows.

Standardizing on a transparent, Apache-licensed VAE supports auditability and compliance requirements, ensures consistent reconstruction quality across internal assets, and allows future models trained for the same latent space to function as drop-in replacements.

This transparency also enables downstream customization such as lightweight fine-tuning for brand styles or internal visual templates—even for organizations that do not specialize in media but rely on consistent, controllable image generation for marketing materials, product imagery, documentation, or stock-style visuals.

The announcement positions FLUX.2 as an evolution of the FLUX.1 family, with an emphasis on reliability, controllability, and integration into existing creative pipelines rather than one-off demos.

A Shift Toward Production-Centric Image Models

FLUX.2 extends the prior FLUX.1 architecture with more consistent character, layout, and style adherence across up to ten reference images.

The system maintains coherence at 4-megapixel resolutions for both generation and editing tasks, enabling use cases such as product visualization, brand-aligned asset creation, and structured design workflows.

The model also improves prompt following across multi-part instructions while reducing failure modes related to lighting, spatial logic, and world knowledge.

In parallel, Black Forest Labs continues to follow an open-core release strategy. The company provides hosted, performance-optimized versions of FLUX.2 for commercial deployments, while also publishing inspectable open-weight models that researchers and independent developers can run locally. This approach extends a track record begun with FLUX.1, which became the most widely used open image model globally.

Model Variants and Deployment Options

Flux.2 arrives with 5 variants as follows:

• Flux.2 [Pro]: This is the highest-performance tier, intended for applications that require minimal latency and maximal visual fidelity. It is available through the BFL Playground, the FLUX API, and partner platforms. The model aims to match leading closed-weight systems in prompt adherence and image quality while reducing compute demand.

• Flux.2 [Flex]: This version exposes parameters such as the number of sampling steps and the guidance scale. The design enables developers to tune the trade-offs between speed, text accuracy, and detail fidelity. In practice, this enables workflows where low-step previews can be generated quickly before higher-step renders are invoked.

• Flux.2 [Dev]: The most notable release for the open ecosystem is the 32-billion-parameter open-weight checkpoint which integrates text-to-image generation and image editing into a single model. It supports multi-reference conditioning without requiring separate modules or pipelines. The model can run locally using BFL’s reference inference code or optimized fp8 implementations developed in partnership with NVIDIA and ComfyUI. Hosted inference is also available via FAL, Replicate, Runware, Verda, TogetherAI, Cloudflare, and DeepInfra.

• Flux.2 [Klein]: Coming soon, this size-distilled model is released under Apache 2.0 and is intended to offer improved performance relative to comparable models of the same size trained from scratch. A beta program is currently open.

• Flux.2 – VAE: Released under the enterprise friendly (even for commercial use) Apache 2.0 license, updated variational autoencoder provides the latent space that underpins all Flux.2 variants. The VAE emphasizes an optimized balance between reconstruction fidelity, learnability, and compression rate—a long-standing challenge for latent-space generative architectures.

Benchmark Performance

Black Forest Labs published two sets of evaluations highlighting FLUX.2’s performance relative to other open-weight and hosted image-generation models. In head-to-head win-rate comparisons across three categories—text-to-image generation, single-reference editing, and multi-reference editing—FLUX.2 [Dev] led all open-weight alternatives by a substantial margin.

It achieved a 66.6% win rate in text-to-image generation (vs. 51.3% for Qwen-Image and 48.1% for Hunyuan Image 3.0), 59.8% in single-reference editing (vs. 49.3% for Qwen-Image and 41.2% for FLUX.1 Kontext), and 63.6% in multi-reference editing (vs. 36.4% for Qwen-Image). These results reflect consistent gains over both earlier FLUX.1 models and contemporary open-weight systems.

A second benchmark compared model quality using ELO scores against approximate per-image cost. In this analysis, FLUX.2 [Pro], FLUX.2 [Flex], and FLUX.2 [Dev] cluster in the upper-quality, lower-cost region of the chart, with ELO scores in the ~1030–1050 band while operating in the 2–6 cent range.

By contrast, earlier models such as FLUX.1 Kontext [max] and Hunyuan Image 3.0 appear significantly lower on the ELO axis despite similar or higher per-image costs. Only proprietary competitors like Nano Banana 2 reach higher ELO levels, but at noticeably elevated cost. According to BFL, this positions FLUX.2’s variants as offering strong quality–cost efficiency across performance tiers, with FLUX.2 [Dev] in particular delivering near–top-tier quality while remaining one of the lowest-cost options in its class.

Pricing via API and Comparison to Nano Banana Pro

A pricing calculator on BFL’s site indicates that FLUX.2 [Pro] is billed at roughly $0.03 per megapixel of combined input and output. A standard 1024×1024 (1 MP) generation costs $0.030, and higher resolutions scale proportionally. The calculator also counts input images toward total megapixels, suggesting that multi-image reference workflows will have higher per-call costs.

By contrast, Google’s Gemini 3 Pro Image Preview aka "Nano Banana Pro," currently prices image output at $120 per 1M tokens, resulting in a cost of $0.134 per 1K–2K image (up to 2048×2048) and $0.24 per 4K image. Image input is billed at $0.0011 per image, which is negligible compared to output costs.

While Gemini’s model uses token-based billing, its effective per-image pricing places 1K–2K images at more than 4× the cost of a 1 MP FLUX.2 [Pro] generation, and 4K outputs at roughly 8× the cost of a similar-resolution FLUX.2 output if scaled proportionally.

In practical terms, the available data suggests that FLUX.2 [Pro] currently offers significantly lower per-image pricing, particularly for high-resolution outputs or multi-image editing workflows, whereas Gemini 3 Pro’s preview tier is positioned as a higher-cost, token-metered service with more variability depending on resolution.

Technical Design and the Latent Space Overhaul

FLUX.2 is built on a latent flow matching architecture, combining a rectified flow transformer with a vision-language model based on Mistral-3 (24B). The VLM contributes semantic grounding and contextual understanding, while the transformer handles spatial structure, material representation, and lighting behavior.

A major component of the update is the re-training of the model’s latent space. The FLUX.2 VAE integrates advances in semantic alignment, reconstruction quality, and representational learnability drawn from recent research on autoencoder optimization. Earlier models often faced trade-offs in the learnability–quality–compression triad: highly compressed spaces increase training efficiency but degrade reconstructions, while wider bottlenecks can reduce the ability of generative models to learn consistent transformations.

According to BFL’s research data, the FLUX.2 VAE achieves lower LPIPS distortion than the FLUX.1 and SD autoencoders while also improving generative FID. This balance allows FLUX.2 to support high-fidelity editing—an area that typically demands reconstruction accuracy—and still maintain competitive learnability for large-scale generative training.

Capabilities Across Creative Workflows

The most significant functional upgrade is multi-reference support. FLUX.2 can ingest up to ten reference images and maintain identity, product details, or stylistic elements across the output. This feature is relevant for commercial applications such as merchandising, virtual photography, storyboarding, and branded campaign development.

The system’s typography improvements address a persistent challenge for diffusion- and flow-based architectures. FLUX.2 is able to generate legible fine text, structured layouts, UI elements, and infographic-style assets with greater reliability. This capability, combined with flexible aspect ratios and high-resolution editing, broadens the use cases where text and image jointly define the final output.

FLUX.2 enhances instruction following for multi-step, compositional prompts, enabling more predictable outcomes in constrained workflows. The model exhibits better grounding in physical attributes—such as lighting and material behavior—reducing inconsistencies in scenes requiring photoreal equilibrium.

Ecosystem and Open-Core Strategy

Black Forest Labs continues to position its models within an ecosystem that blends open research with commercial reliability. The FLUX.1 open models helped establish the company’s reach across both the developer and enterprise markets, and FLUX.2 expands this structure: tightly optimized commercial endpoints for production deployments and open, composable checkpoints for research and community experimentation.

The company emphasizes transparency through published inference code, open-weight VAE release, prompting guides, and detailed architectural documentation. It also continues to recruit talent in Freiburg and San Francisco as it pursues a longer-term roadmap toward multimodal models that unify perception, memory, reasoning, and generation.

Background: Flux and the Formation of Black Forest Labs

Black Forest Labs (BFL) was founded in 2024 by Robin Rombach, Patrick Esser, and Andreas Blattmann, the original creators of Stable Diffusion. Their move from Stability AI came at a moment of turbulence for the broader open-source generative AI community, and the launch of BFL signaled a renewed effort to build accessible, high-performance image models. The company secured $31 million in seed funding led by Andreessen Horowitz, with additional support from Brendan Iribe, Michael Ovitz, and Garry Tan, providing early validation for its technical direction.

BFL’s first major release, FLUX.1, introduced a 12-billion-parameter architecture available in Pro, Dev, and Schnell variants. It quickly gained a reputation for output quality that matched or exceeded closed-source competitors such as Midjourney v6 and DALL·E 3, while the Dev and Schnell versions reinforced the company’s commitment to open distribution. FLUX.1 also saw rapid adoption in downstream products, including xAI’s Grok 2, and arrived amid ongoing industry discussions about dataset transparency, responsible model usage, and the role of open-source distribution. BFL published strict usage policies aimed at preventing misuse and non-consensual content generation.

In late 2024, BFL expanded the lineup with Flux 1.1 Pro, a proprietary high-speed model delivering sixfold generation speed improvements and achieving leading ELO scores on Artificial Analysis. The company launched a paid API alongside the release, enabling configurable integrations with adjustable resolution, model choice, and moderation settings at pricing that began at $0.04 per image.

Partnerships with TogetherAI, Replicate, FAL, and Freepik broadened access and made the model available to users without the need for self-hosting, extending BFL’s reach across commercial and creator-oriented platforms.

These developments unfolded against a backdrop of accelerating competition in generative media.

Implications for Enterprise Technical Decision Makers

The FLUX.2 release carries distinct operational implications for enterprise teams responsible for AI engineering, orchestration, data management, and security. For AI engineers responsible for model lifecycle management, the availability of both hosted endpoints and open-weight checkpoints enables flexible integration paths.

FLUX.2’s multi-reference capabilities and expanded resolution support reduce the need for bespoke fine-tuning pipelines when handling brand-specific or identity-consistent outputs, lowering development overhead and accelerating deployment timelines. The model’s improved prompt adherence and typography performance also reduce iterative prompting cycles, which can have a measurable impact on production workload efficiency.

Teams focused on AI orchestration and operational scaling benefit from the structure of FLUX.2’s product family. The Pro tier offers predictable latency characteristics suitable for pipeline-critical workloads, while the Flex tier enables direct control over sampling steps and guidance parameters, aligning with environments that require strict performance tuning.

Open-weight access for the Dev model facilitates the creation of custom containerized deployments and allows orchestration platforms to manage the model under existing CI/CD practices. This is particularly relevant for organizations balancing cutting-edge tooling with budget constraints, as self-hosted deployments offer cost control at the expense of in-house optimization requirements.

Data engineering stakeholders gain advantages from the model’s latent architecture and improved reconstruction fidelity. High-quality, predictable image representations reduce downstream data-cleaning burdens in workflows where generated assets feed into analytics systems, creative automation pipelines, or multimodal model development.

Because FLUX.2 consolidates text-to-image and image-editing functions into a single model, it simplifies integration points and reduces the complexity of data flows across storage, versioning, and monitoring layers. For teams managing large volumes of reference imagery, the ability to incorporate up to ten inputs per generation may also streamline asset management processes by shifting more variation handling into the model rather than external tooling.

For security teams, FLUX.2’s open-core approach introduces considerations related to access control, model governance, and API usage monitoring. Hosted FLUX.2 endpoints allow for centralized enforcement of security policies and reduce local exposure to model weights, which may be preferable for organizations with stricter compliance requirements.

Conversely, open-weight deployments require internal controls for model integrity, version tracking, and inference-time monitoring to prevent misuse or unapproved modifications. The model’s handling of typography and realistic compositions also reinforces the need for established content governance frameworks, particularly where generative systems interface with public-facing channels.

Across these roles, FLUX.2’s design emphasizes predictable performance characteristics, modular deployment options, and reduced operational friction. For enterprises with lean teams or rapidly evolving requirements, the release offers a set of capabilities aligned with practical constraints around speed, quality, budget, and model governance.

FLUX.2 marks a substantial iterative improvement in Black Forest Labs’ generative image stack, with notable gains in multi-reference consistency, text rendering, latent space quality, and structured prompt adherence. By pairing fully managed offerings with open-weight checkpoints, BFL maintains its open-core model while extending its relevance to commercial creative workflows. The release demonstrates a shift from experimental image generation toward more predictable, scalable, and controllable systems suited for operational use.

OpenAI expanded its data residency regions for ChatGPT and its API, giving enterprise users the option to store and process their data closest to their business operations and better comply with local regulations. This expansion removes one of the biggest compliance blockers preventing global enterprises from deploying ChatGPT at scale.

Data residency, often an overlooked piece of the enterprise AI puzzle, processes and governs data according to the laws and customs of the countries where it is stored. 

ChatGPT Enterprise and Edu subscribers can now choose to have their data processed in: 

• Europe (European Economic Area and Switzerland)

• United Kingdom

• United States

• Canada

• Japan

• South Korea

• Singapore

• India

• Australia

• United Arab Emirates

OpenAI said in a blog post that it “plans to expand availability to additional regions over time.” 

Customers can store data such as conversations, uploaded files, custom GPTs, and image-generation artifacts. This applies only to data at rest, not while it moves through a system or when it is used for inference. OpenAI’s documentation notes that, for now, inference residency remains available only in the U.S.  

ChatGPT Enterprise and Edu users can set up new workspaces with data residency. Enterprise customers on the API who have been approved for advanced data controls can enable data residency by creating a new project and selecting their preferred region.

OpenAI first began offering data residency in Europe in February this year. The European Union has some of the strictest data regulations globally, based on the GDPR. 

The importance of data residency

Enterprises until now had fewer choices for processing data flowing through ChatGPT. For example, some organizational data would be processed under U.S. law rather than under European rules. 

Enterprises risk violating data compliance rules if their data at rest is processed elsewhere and does not meet strict policies. 

“With over 1 million business customers around the world directly using OpenAI, we have expanded where we offer data residency — allowing business customers to store data in certain regions, helping organizations meet local regulatory and data protection requirements,” the company said in its blog post. 

However, enterprises must also understand that if they are using a connector or integration within ChatGPT, those applications have different data residency rules. When OpenAI launched company knowledge for ChatGPT, it warned users that depending on the connector they use, data residency may be limited to the U.S. 

President Donald Trump’s new “Genesis Mission” unveiled Monday is billed as a generational leap in how the United States does science akin to the Manhattan Project that created the atomic bomb during World War II.

The executive order directs the Department of Energy (DOE) to build a “closed-loop AI experimentation platform” that links the country’s 17 national laboratories, federal supercomputers, and decades of government scientific data into “one cooperative system for research.”

The White House fact sheet casts the initiative as a way to “transform how scientific research is conducted” and “accelerate the speed of scientific discovery,” with priorities spanning biotechnology, critical materials, nuclear fission and fusion, quantum information science, and semiconductors.

DOE’s own release calls it “the world’s most complex and powerful scientific instrument ever built” and quotes Under Secretary for Science Darío Gil describing it as a “closed-loop system” linking the nation’s most advanced facilities, data, and computing into “an engine for discovery that doubles R&D productivity.”

What the administration has not provided is just as striking: no public cost estimate, no explicit appropriation, and no breakdown of who will pay for what. Major news outlets including Reuters, Associated Press, Politico, and others have all noted that the order “does not specify new spending or a budget request,” or that funding will depend on future appropriations and previously passed legislation.

That omission, combined with the initiative’s scope and timing, raises questions not only about how Genesis will be funded and to what extent, but about who it might quietly benefit.

“So is this just a subsidy for big labs or what?”

Soon after DOE promoted the mission on X, Teknium of the small U.S. AI lab Nous Research posted a blunt reaction: “So is this just a subsidy for big labs or what.”

The line has become a shorthand for a growing concern in the AI community: that the U.S. government could offer some sort of public subsidy for large AI firms facing staggering and rising compute and data costs.

That concern is grounded in recent, well-sourced reporting on OpenAI’s finances and infrastructure commitments. Documents obtained and analyzed by tech public relations professional and AI critic Ed Zitron describe a cost structure that has exploded as the company has scaled models like GPT-4, GPT-4.1, and GPT-5.1.

The Register has separately inferred from Microsoft quarterly earnings statements that OpenAI lost about $13.5 billion on $4.3 billion in revenue in the first half of 2025 alone. Other outlets and analysts have highlighted projections that show tens of billions in annual losses later this decade if spending and revenue follow current trajectories

By contrast, Google DeepMind trained its recent Gemini 3 flagship LLM on the company’s own TPU hardware and in its own data centers, giving it a structural advantage in cost per training run and energy management, as covered in Google’s own technical blogs and subsequent financial reporting.

Viewed against that backdrop, an ambitious federal project that promises to integrate “world-class supercomputers and datasets into a unified, closed-loop AI platform” and “power robotic laboratories” sounds, to some observers, like more than a pure science accelerator. It could, depending on how access is structured, also ease the capital bottlenecks facing private frontier-model labs.

The executive order explicitly anticipates partnerships with “external partners possessing advanced AI, data, or computing capabilities,” to be governed through cooperative research and development agreements, user-facility partnerships, and data-use and model-sharing agreements. That category clearly includes firms like OpenAI, Anthropic, Google, and other major AI players—even if none are named.

What the order does not do is guarantee those companies access, spell out subsidized pricing, or earmark public money for their training runs. Any claim that OpenAI, Anthropic, or Google “just got access” to federal supercomputing or national-lab data is, at this point, an interpretation of how the framework could be used, not something the text actually promises.

Furthermore, the executive order makes no mention of open-source model development — an omission that stands out in light of remarks last year from Vice President JD Vance, when, prior to assuming office and while serving as a Senator from Ohio and participating in a hearing, he warned against regulations designed to protect incumbent tech firms and was widely praised by open-source advocates.

Closed-loop discovery and “autonomous scientific agents”

Another viral reaction came from AI influencer Chris (@chatgpt21 on X), who wrote in an X post that that OpenAI, Anthropic, and Google have already “got access to petabytes of proprietary data” from national labs, and that DOE labs have been “hoarding experimental data for decades.” The public record supports a narrower claim.

The order and fact sheet describe “federal scientific datasets—the world’s largest collection of such datasets, developed over decades of Federal investments” and direct agencies to identify data that can be integrated into the platform “to the extent permitted by law.”

DOE’s announcement similarly talks about unleashing “the full power of our National Laboratories, supercomputers, and data resources.”

It is true that the national labs hold enormous troves of experimental data. Some of it is already public via the Office of Scientific and Technical Information (OSTI) and other repositories; some is classified or export-controlled; much is under-used because it sits in fragmented formats and systems. But there is no public document so far that states private AI companies have now been granted blanket access to this data, or that DOE characterizes past practice as “hoarding.”

What is clear is that the administration wants to unlock more of this data for AI-driven research and to do so in coordination with external partners. Section 5 of the order instructs DOE and the Assistant to the President for Science and Technology to create standardized partnership frameworks, define IP and licensing rules, and set “stringent data access and management processes and cybersecurity standards for non-Federal collaborators accessing datasets, models, and computing environments.”

A moonshot with an open question at the center

Taken at face value, the Genesis Mission is an ambitious attempt to use AI and high-performance computing to speed up everything from fusion research to materials discovery and pediatric cancer work, using decades of taxpayer-funded data and instruments that already exist inside the federal system. The executive order spends considerable space on governance: coordination through the National Science and Technology Council, new fellowship programs, and annual reporting on platform status, integration progress, partnerships, and scientific outcomes.

Yet the initiative also lands at a moment when frontline AI labs are buckling under their own compute bills, when one of them—OpenAI—is reported to be spending more on running models than it earns in revenue, and when investors are openly debating whether the current business model for proprietary frontier AI is sustainable without some form of outside support.

In that environment, a federally funded, closed-loop AI discovery platform that centralizes the country’s most powerful supercomputers and data is inevitably going to be read in more than one way. It may become a genuine engine for public science. It may also become a crucial piece of infrastructure for the very companies driving today’s AI arms race.

For now, one fact is undeniable: the administration has launched a mission it compares to the Manhattan Project without telling the public what it will cost, how the money will flow, or exactly who will be allowed to plug into it.

How enterprise tech leaders should interpret the Genesis Mission

For enterprise teams already building or scaling AI systems, the Genesis Mission signals a shift in how national infrastructure, data governance, and high-performance compute will evolve in the U.S.—and those signals matter even before the government publishes a budget.

The initiative outlines a federated, AI-driven scientific ecosystem where supercomputers, datasets, and automated experimentation loops operate as tightly integrated pipelines.

That direction mirrors the trajectory many companies are already moving toward: larger models, more experimentation, heavier orchestration, and a growing need for systems that can manage complex workloads with reliability and traceability.

Even though Genesis is aimed at science, its architecture hints at what will become expected norms across American industries.

The lack of cost detail around Genesis does not directly alter enterprise roadmaps, but it does reinforce the broader reality that compute scarcity, escalating cloud costs, and rising standards for AI model governance will remain central challenges.

Companies that already struggle with constrained budgets or tight headcount—particularly those responsible for deployment pipelines, data integrity, or AI security—should view Genesis as early confirmation that efficiency, observability, and modular AI infrastructure will remain essential.

As the federal government formalizes frameworks for data access, experiment traceability, and AI agent oversight, enterprises may find that future compliance regimes or partnership expectations take cues from these federal standards.

Genesis also underscores the growing importance of unifying data sources and ensuring that models can operate across diverse, sometimes sensitive environments. Whether managing pipelines across multiple clouds, fine-tuning models with domain-specific datasets, or securing inference endpoints, enterprise technical leaders will likely see increased pressure to harden systems, standardize interfaces, and invest in complex orchestration that can scale safely.

The mission’s emphasis on automation, robotic workflows, and closed-loop model refinement may shape how enterprises structure their internal AI R&D, encouraging them to adopt more repeatable, automated, and governable approaches to experimentation.

Here is what enterprise leaders should be doing now:

1. Expect increased federal involvement in AI infrastructure and data governance. This may indirectly shape cloud availability, interoperability standards, and model-governance expectations.

2. Track “closed-loop” AI experimentation models. This may preview future enterprise R&D workflows and reshape how ML teams build automated pipelines.

3. Prepare for rising compute costs and consider efficiency strategies. This includes smaller models, retrieval-augmented systems, and mixed-precision training.

4. Strengthen AI-specific security practices. Genesis signals that the federal government is escalating expectations for AI system integrity and controlled access.

5. Plan for potential public–private interoperability standards. Enterprises that align early may gain a competitive edge in partnerships and procurement.

Overall, Genesis does not change day-to-day enterprise AI operations today. But it strongly signals where federal and scientific AI infrastructure is heading—and that direction will inevitably influence the expectations, constraints, and opportunities enterprises face as they scale their own AI capabilities.

Anthropic released its most capable artificial intelligence model yet on Monday, slashing prices by roughly two-thirds while claiming state-of-the-art performance on software engineering tasks — a strategic move that intensifies the AI startup's competition with deep-pocketed rivals OpenAI and Google.

The new model, Claude Opus 4.5, scored higher on Anthropic's most challenging internal engineering assessment than any human job candidate in the company's history, according to materials reviewed by VentureBeat. The result underscores both the rapidly advancing capabilities of AI systems and growing questions about how the technology will reshape white-collar professions.

The Amazon-backed company is pricing Claude Opus 4.5 at $5 per million input tokens and $25 per million output tokens — a dramatic reduction from the $15 and $75 rates for its predecessor, Claude Opus 4.1, released earlier this year. The move makes frontier AI capabilities accessible to a broader swath of developers and enterprises while putting pressure on competitors to match both performance and pricing.

"We want to make sure this really works for people who want to work with these models," said Alex Albert, Anthropic's head of developer relations, in an exclusive interview with VentureBeat. "That is really our focus: How can we enable Claude to be better at helping you do the things that you don't necessarily want to do in your job?"

The announcement comes as Anthropic races to maintain its position in an increasingly crowded field. OpenAI recently released GPT-5.1 and a specialized coding model called Codex Max that can work autonomously for extended periods. Google unveiled Gemini 3 just last week, prompting concerns even from OpenAI about the search giant's progress, according to a recent report from The Information.

Opus 4.5 demonstrates improved judgment on real-world tasks, developers say

Anthropic's internal testing revealed what the company describes as a qualitative leap in Claude Opus 4.5's reasoning capabilities. The model achieved 80.9% accuracy on SWE-bench Verified, a benchmark measuring real-world software engineering tasks, outperforming OpenAI's GPT-5.1-Codex-Max (77.9%), Anthropic's own Sonnet 4.5 (77.2%), and Google's Gemini 3 Pro (76.2%), according to the company's data. The result marks a notable advance over OpenAI's current state-of-the-art model, which was released just five days earlier.

But the technical benchmarks tell only part of the story. Albert said employee testers consistently reported that the model demonstrates improved judgment and intuition across diverse tasks — a shift he described as the model developing a sense of what matters in real-world contexts.

"The model just kind of gets it," Albert said. "It just has developed this sort of intuition and judgment on a lot of real world things that feels qualitatively like a big jump up from past models."

He pointed to his own workflow as an example. Previously, Albert said, he would ask AI models to gather information but hesitated to trust their synthesis or prioritization. With Opus 4.5, he's delegating more complete tasks, connecting it to Slack and internal documents to produce coherent summaries that match his priorities.

Opus 4.5 outscores all human candidates on company's toughest engineering test

The model's performance on Anthropic's internal engineering assessment marks a notable milestone. The take-home exam, designed for prospective performance engineering candidates, is meant to evaluate technical ability and judgment under time pressure within a prescribed two-hour limit.

Using a technique called parallel test-time compute — which aggregates multiple attempts from the model and selects the best result — Opus 4.5 scored higher than any human candidate who has taken the test, according to company. Without a time limit, the model matched the performance of the best-ever human candidate when used within Claude Code, Anthropic's coding environment.

The company acknowledged that the test doesn't measure other crucial professional skills such as collaboration, communication, or the instincts that develop over years of experience. Still, Anthropic said the result "raises questions about how AI will change engineering as a profession."

Albert emphasized the significance of the finding. "I think this is kind of a sign, maybe, of what's to come around how useful these models can actually be in a work context and for our jobs," he said. "Of course, this was an engineering task, and I would say models are relatively ahead in engineering compared to other fields, but I think it's a really important signal to pay attention to."

Dramatic efficiency improvements cut token usage by up to 76% on key benchmarks

Beyond raw performance, Anthropic is betting that efficiency improvements will differentiate Claude Opus 4.5 in the market. The company says the model uses dramatically fewer tokens — the units of text that AI systems process — to achieve similar or better outcomes compared to predecessors.

At a medium effort level, Opus 4.5 matches the previous Sonnet 4.5 model's best score on SWE-bench Verified while using 76% fewer output tokens, according to Anthropic. At the highest effort level, Opus 4.5 exceeds Sonnet 4.5 performance by 4.3 percentage points while still using 48% fewer tokens.

To give developers more control, Anthropic introduced an "effort parameter" that allows users to adjust how much computational work the model applies to each task — balancing performance against latency and cost.

Enterprise customers provided early validation of the efficiency claims. "Opus 4.5 beats Sonnet 4.5 and competition on our internal benchmarks, using fewer tokens to solve the same problems," said Michele Catasta, president of Replit, a cloud-based coding platform, in a statement to VentureBeat. "At scale, that efficiency compounds."

GitHub's chief product officer, Mario Rodriguez, said early testing shows Opus 4.5 "surpasses internal coding benchmarks while cutting token usage in half, and is especially well-suited for tasks like code migration and code refactoring."

Early customers report AI agents that learn from experience and refine their own skills

One of the most striking capabilities demonstrated by early customers involves what Anthropic calls "self-improving agents" — AI systems that can refine their own performance through iterative learning.

Rakuten, the Japanese e-commerce and internet company, tested Claude Opus 4.5 on automation of office tasks. "Our agents were able to autonomously refine their own capabilities — achieving peak performance in 4 iterations while other models couldn't match that quality after 10," said Yusuke Kaji, Rakuten's general manager of AI for business.

Albert explained that the model isn't updating its own weights — the fundamental parameters that define an AI system's behavior — but rather iteratively improving the tools and approaches it uses to solve problems. "It was iteratively refining a skill for a task and seeing that it's trying to optimize the skill to get better performance so it could accomplish this task," he said.

The capability extends beyond coding. Albert said Anthropic has observed significant improvements in creating professional documents, spreadsheets, and presentations. "They're saying that this has been the biggest jump they've seen between model generations," Albert said. "So going even from Sonnet 4.5 to Opus 4.5, bigger jump than any two models back to back in the past."

Fundamental Research Labs, a financial modeling firm, reported that "accuracy on our internal evals improved 20%, efficiency rose 15%, and complex tasks that once seemed out of reach became achievable," according to co-founder Nico Christie.

New features target Excel users, Chrome workflows and eliminate chat length limits

Alongside the model release, Anthropic rolled out a suite of product updates aimed at enterprise users. Claude for Excel became generally available for Max, Team, and Enterprise users with new support for pivot tables, charts, and file uploads. The Chrome browser extension is now available to all Max users.

Perhaps most significantly, Anthropic introduced "infinite chats" — a feature that eliminates context window limitations by automatically summarizing earlier parts of conversations as they grow longer. "Within Claude AI, within the product itself, you effectively get this kind of infinite context window due to the compaction, plus some memory things that we're doing," Albert explained.

For developers, Anthropic released "programmatic tool calling," which allows Claude to write and execute code that invokes functions directly. Claude Code gained an updated "Plan Mode" and became available on desktop in research preview, enabling developers to run multiple AI agent sessions in parallel.

Market heats up as OpenAI, Google race to match performance and pricing

Anthropic reached $2 billion in annualized revenue during the first quarter of 2025, more than doubling from $1 billion in the prior period. The number of customers spending more than $100,000 annually jumped eightfold year-over-year.

The rapid release of Opus 4.5 — just weeks after Haiku 4.5 in October and Sonnet 4.5 in September — reflects broader industry dynamics. OpenAI released multiple GPT-5 variants throughout 2025, including a specialized Codex Max model in November that can work autonomously for up to 24 hours. Google shipped Gemini 3 in mid-November after months of development.

Albert attributed Anthropic's accelerated pace partly to using Claude to speed its own development. "We're seeing a lot of assistance and speed-up by Claude itself, whether it's on the actual product building side or on the model research side," he said.

The pricing reduction for Opus 4.5 could pressure margins while potentially expanding the addressable market. "I'm expecting to see a lot of startups start to incorporate this into their products much more and feature it prominently," Albert said.

Yet profitability remains elusive for leading AI labs as they invest heavily in computing infrastructure and research talent. The AI market is projected to top $1 trillion in revenue within a decade, but no single provider has established dominant market position—even as models reach a threshold where they can meaningfully automate complex knowledge work.

Michael Truell, CEO of Cursor, an AI-powered code editor, called Opus 4.5 "a notable improvement over the prior Claude models inside Cursor, with improved pricing and intelligence on difficult coding tasks." Scott Wu, CEO of Cognition, an AI coding startup, said the model delivers "stronger results on our hardest evaluations and consistent performance through 30-minute autonomous coding sessions."

For enterprises and developers, the competition translates to rapidly improving capabilities at falling prices. But as AI performance on technical tasks approaches—and sometimes exceeds—human expert levels, the technology's impact on professional work becomes less theoretical.

When asked about the engineering exam results and what they signal about AI's trajectory, Albert was direct: "I think it's a really important signal to pay attention to."

Microsoft has introduced Fara-7B, a new 7-billion parameter model designed to act as a Computer Use Agent (CUA) capable of performing complex tasks directly on a user’s device. Fara-7B sets new state-of-the-art results for its size, providing a way to build AI agents that don’t rely on massive, cloud-dependent models and can run on compact systems with lower latency and enhanced privacy.

While the model is an experimental release, its architecture addresses a primary barrier to enterprise adoption: data security. Because Fara-7B is small enough to run locally, it allows users to automate sensitive workflows, such as managing internal accounts or processing sensitive company data, without that information ever leaving the device. 

How Fara-7B sees the web

Fara-7B is designed to navigate user interfaces using the same tools a human does: a mouse and keyboard. The model operates by visually perceiving a web page through screenshots and predicting specific coordinates for actions like clicking, typing, and scrolling.

Crucially, Fara-7B does not rely on "accessibility trees,” the underlying code structure that browsers use to describe web pages to screen readers. Instead, it relies solely on pixel-level visual data. This approach allows the agent to interact with websites even when the underlying code is obfuscated or complex.

According to Yash Lara, Senior PM Lead at Microsoft Research, processing all visual input on-device creates true "pixel sovereignty," since screenshots and the reasoning needed for automation remain on the user’s device. "This approach helps organizations meet strict requirements in regulated sectors, including HIPAA and GLBA," he told VentureBeat in written comments.

In benchmarking tests, this visual-first approach has yielded strong results. On WebVoyager, a standard benchmark for web agents, Fara-7B achieved a task success rate of 73.5%. This outperforms larger, more resource-intensive systems, including GPT-4o, when prompted to act as a computer use agent (65.1%) and the native UI-TARS-1.5-7B model (66.4%).

Efficiency is another key differentiator. In comparative tests, Fara-7B completed tasks in approximately 16 steps on average, compared to roughly 41 steps for the UI-TARS-1.5-7B model.

Handling risks

The transition to autonomous agents is not without risks, however. Microsoft notes that Fara-7B shares limitations common to other AI models, including potential hallucinations, mistakes in following complex instructions, and accuracy degradation on intricate tasks.

To mitigate these risks, the model was trained to recognize "Critical Points." A Critical Point is defined as any situation requiring a user's personal data or consent before an irreversible action occurs, such as sending an email or completing a financial transaction. Upon reaching such a juncture, Fara-7B is designed to pause and explicitly request user approval before proceeding. 

Managing this interaction without frustrating the user is a key design challenge. "Balancing robust safeguards such as Critical Points with seamless user journeys is key," Lara said. "Having a UI, like Microsoft Research’s Magentic-UI, is vital for giving users opportunities to intervene when necessary, while also helping to avoid approval fatigue." Magentic-UI is a research prototype designed specifically to facilitate these human-agent interactions. Fara-7B is designed to run in Magentic-UI.

Distilling complexity into a single model

The development of Fara-7B highlights a growing trend in knowledge distillation, where the capabilities of a complex system are compressed into a smaller, more efficient model.

Creating a CUA usually requires massive amounts of training data showing how to navigate the web. Collecting this data via human annotation is prohibitively expensive. To solve this, Microsoft used a synthetic data pipeline built on Magentic-One, a multi-agent framework. In this setup, an "Orchestrator" agent created plans and directed a "WebSurfer" agent to browse the web, generating 145,000 successful task trajectories.

The researchers then "distilled" this complex interaction data into Fara-7B, which is built on Qwen2.5-VL-7B, a base model chosen for its long context window (up to 128,000 tokens) and its strong ability to connect text instructions to visual elements on a screen. While the data generation required a heavy multi-agent system, Fara-7B itself is a single model, showing that a small model can effectively learn advanced behaviors without needing complex scaffolding at runtime.

The training process relied on supervised fine-tuning, where the model learns by mimicking the successful examples generated by the synthetic pipeline.

Looking forward

While the current version was trained on static datasets, future iterations will focus on making the model smarter, not necessarily bigger. "Moving forward, we’ll strive to maintain the small size of our models," Lara said. "Our ongoing research is focused on making agentic models smarter and safer, not just larger." This includes exploring techniques like reinforcement learning (RL) in live, sandboxed environments, which would allow the model to learn from trial and error in real-time.

Microsoft has made the model available on Hugging Face and Microsoft Foundry under an MIT license. However, Lara cautions that while the license allows for commercial use, the model is not yet production-ready. "You can freely experiment and prototype with Fara‑7B under the MIT license," he says, "but it’s best suited for pilots and proofs‑of‑concept rather than mission‑critical deployments."

Remember the first time you heard your company was going AI-first?

Maybe it came through an all-hands that felt different from the others. The CEO said, “By Q3, every team should have integrated AI into their core workflows,” and the energy in the room (or on the Zoom) shifted. You saw a mix of excitement and anxiety ripple through the crowd.

Maybe you were one of the curious ones. Maybe you’d already built a Python script that summarized customer feedback, saving your team three hours every week. Or maybe you’d stayed late one night just to see what would happen if you combined a dataset with a large language model (LLM) prompt. Maybe you’re one of those who’d already let curiosity lead you somewhere unexpected.

But this announcement felt different because suddenly, what had been a quiet act of curiosity was now a line in a corporate OKR. Maybe you didn’t know it yet, but something fundamental had shifted in how innovation would happen inside your company.

How innovation happens

Real transformation rarely looks like the PowerPoint version, and almost never follows the org chart.

Think about the last time something genuinely useful spread at work. It wasn't because of a vendor pitch or a strategic initiative, was it? More likely, someone stayed late one night, when no one was watching, found something that cut hours of busywork, and mentioned it at lunch the next day. “Hey, try this.” They shared it in a Slack thread and, in a week, half the team was using it.

The developer who used GPT to debug code wasn’t trying to make a strategic impact. She just needed to get home earlier to her kids. The ops manager who automated his spreadsheet didn’t need permission. He just needed more sleep.

This is the invisible architecture of progress — these informal networks where curiosity flows like water through concrete… finding every crack, every opening.

But watch what happens when leadership notices. What used to be effortless and organic becomes mandated. And the thing that once worked because it was free suddenly stops being as effective the moment it’s measured.

The great reversal

It usually begins quietly. Often when a competitor announces new AI features, — like AI-powered onboarding or end-to-end support automation — claiming 40% efficiency gains.

The next morning, your CEO calls an emergency meeting. The room gets still. Someone clears their throat. And you can feel everyone doing mental math about their job security. “If they’re that far ahead, what does that mean for us?”

That afternoon, your company has a new priority. Your CEO says, “We need an AI strategy. Yesterday.”

Here's how that message usually ripples down the org chart:

• At the C-suite: “We need an AI strategy to stay competitive.”

• At the VP level: “Every team needs an AI initiative.”

• At the manager level: “We need a plan by Friday.”

• At your level: “I just need to find something that looks like AI.”

Each translation adds pressure while subtracting understanding. Everyone still cares, but that translation changes intent. What begins as a question worth asking becomes a script everyone follows blindly.

Eventually, the performance of innovation replaces the thing itself. There’s a strange pressure to look like you’re moving fast, even when you’re not sure where you’re actually going.

This repeats across industries

A competitor declared they’re going AI-first. Another publishes a case study about replacing support with LLMs. And a third shares a graph showing productivity gains. Within days, boardrooms everywhere start echoing the same message: “We should be doing this. Everyone else already is, and we can’t fall behind.”

So the work begins. Then come the task forces, the town halls, the strategy docs and the targets. Teams are asked to contribute initiatives.

But if you’ve been through this before, you know there’s often a difference between what companies announce and what they actually do. Because press releases don’t mention the pilots that stall, or the teams that quietly revert to the old way, or even the tools that get used once and abandoned. You might know someone who was on one of those teams, or you might’ve even been on one yourself.

These aren’t failures of technology or intent. ChatGPT works fine. And teams want to automate their tasks. These failures are organizational, and they happen when we try to imitate outcomes without understanding what created them in the first place.

And so when everyone performs innovation, it becomes almost impossible to tell who’s actually doing it.

Two kinds of leaders

You’ve probably seen both, and it’s very easy to tell which kind you’re working with.

One spends an entire weekend prototyping. They try something new, fail at half of it, and still show up Monday saying, “I built this thing with Claude. It crashed after two hours, but I learned a lot. Wanna see? It's very basic, but it might solve that thing we talked about.”

They try to build understanding. You can tell they’ve actually spent time with AI, and struggled with prompts and hallucinations. Instead of trying to sound certain, they talk about what broke, what almost worked and what they’re still figuring out. They invite you to try something new, because it feels like there’s room to learn. That’s what leading by participation looks like.

The other sends you a directive in Slack: “Leadership wants every team using AI by the end of the quarter. Plans are due by Friday.” They enforce compliance with a decision that's already been made. You can even hear it in their language, and how certain they sound.

The curious leader builds momentum. The performative one builds resentment.

What actually works

You probably don’t need someone to tell you where AI works. You already know because you’ve seen it.

• Customer support: LLMs genuinely help with Tier 1 tickets. They understand intent, draft simple responses and route complexity. Not perfectly, of course, — I’m sure you've seen the failures — but well enough to matter.

• Code assistance: At 2 a.m., when you’re half-delirious and your AI assistant suggests exactly what you need, it feels like having an over-caffeinated junior programmer who never judges your forgotten semicolons. You save minutes at first, then hours, then days.

These small, cumulative wins compound over time. They aren't the impressive transformations promised in decks, but the kind of improvements you can rely on.

But outside these zones, things get murky. AI-driven revops? Fully automated forecasting? You've sat through those demos, and you’ve also seen the enthusiasm fade once the pilot actually begins.

Have the builders of these AI tools failed? Hardly. The technology is evolving, and the products built on top of it are still learning how to walk.

So how can you tell if your company's AI adoption is real? Simple. Just ask someone in finance or ops. Ask what AI tools they use daily. You might get a slight pause or an apologetic smile. “Honestly? Just ChatGPT.” That’s it. Not the $50k enterprise-grade platform from last quarter’s demo or the expensive software suite in the board deck. Just a browser tab, same as any college student writing an essay.

You might make this same confession yourself. Despite all the mandates and initiatives, your most powerful AI tool is probably the same one everyone else uses. So what does this tell us about the gap between what we're supposed to be doing and what we're actually doing?

How to drive change at your company

You've probably discovered this yourself, even if no one's ever put it into words:

1. Model what you mean: Remember that engineering director who screen-shared her messy, live coding session with Cursor? You learned more from watching her debug in real time than from any polished presentation, because vulnerability travels farther than directives.

2. Listen to the edges: You know who's actually using AI effectively in your organization, and they're not always the ones with “AI” in their title. They're the curious ones who've been quietly experimenting, finding what works through trial and error. And that knowledge is worth more than any analyst report.

3. Create permission (not pressure): The people inclined to experiment will always find a way, and the rest won’t be moved by force. The best thing you can do is make the curious feel safe to stay curious.

We're living in this strange moment, caught between the AI that vendors promise and the AI that actually exists on our screens, and it's deeply uncomfortable. The gap between product and promise is wide.

But what I've learned from sitting in that discomfort is that companies that will thrive aren’t the ones that adopted AI first, but the ones that learned through trial and error. They stayed with the discomfort long enough for it to teach them something.

Where will you be six months from now?

By then, your company’s AI-first mandate will have set into motion departmental initiatives, vendor contracts and maybe even some new hires with “AI” in their titles. The dashboards will be green, and the board deck will have a whole slide on AI.

But in the quiet spaces where your actual work happens, what will have meaningfully changed?

Maybe you'll be like the teams that never stopped their quiet experiments. Your customer feedback system might catch the patterns humans miss. Your documentation might update itself. Chances are, if you were building before the mandate, you’ll be building after it fades.

That’s invisible architecture of genuine progress: Patient, and completely uninterested in performance. It doesn't make for great LinkedIn posts, and it resists grand narratives. But it transforms companies in ways that truly last.

Every organization is standing at the same crossroads right now: Look like you’re innovating, or create a culture that fosters real innovation.

The pressure to perform innovation is real, and it’s growing. Most companies will give in and join the theater. But some understand that curiosity can’t be forced, and progress can’t be performed. Because real transformation happens when no one’s watching, in the hands of the people still experimenting, still learning. That’s where the future begins.

Siqi Chen is co-founder and CEO of Runway.

Read more from our guest writers. Or, consider submitting a post of your own! See our guidelines here.