Anthropic releases Fable 5.2 and Mythos 5.2 models with 30% faster speeds

New frontier model release

Anthropic has officially launched Claude Fable 5.2 and Claude Mythos 5.2. These models are designed to handle demanding technical challenges in areas such as software engineering, robotics, and biomedical research. They are available now for paid subscribers and on major cloud platforms like AWS Bedrock and Google Cloud.

The release marks a significant step forward in technical performance for the company. While the standard Claude Opus 5.5 is positioned as a daily driver, these new models serve as industrial-grade engines. They are specifically built for long-horizon autonomous tasks and complex agent swarms.

Developers can access these models through the Claude Platform and various enterprise cloud environments. The rollout includes specific integrations for Claude Code, which defaults to Medium effort. This ensures that high-level reasoning is available to both individual users and large-scale corporate infrastructures.

Speed and pricing updates

The 5.2 architecture introduces a notable 30% increase in standard output generation speed over the previous 5.1 versions. For users who need even faster results, an optional Fast Mode is available to accelerate output tokens by up to 2.5x. These improvements help reduce the time required for complex computations.

Alongside these performance gains, Anthropic has implemented a 20% price cut for developer API usage. The new baseline pricing is set at $8 per million input tokens and $40 per million output tokens. This is a reduction from the previous costs of $10 and $50 respectively.

These changes aim to make high-end AI reasoning more economically viable for developers. By lowering the barrier to entry, Anthropic is positioning Fable 5.2 as a practical tool for heavy-duty industrial applications. The pricing structure supports the deployment of large-scale agentic systems.

Coding and math benchmarks

Claude Fable 5.2 demonstrated significant dominance in coding benchmarks, specifically on FrontierCode 2.0. It achieved a score of 49.1% at Xhigh effort, which is the highest score of any released model on that benchmark. This performance matches the Opus 5.5 model at High effort but costs roughly half the tokens per task.

The model also showed strong results in other engineering and search categories. It scored 82.1% on the WANDR agentic search benchmark. These numbers highlight its ability to handle complex simulations and hardware modeling.

In the realm of formal mathematics, the model showed unique capabilities in a distributed cluster test. A swarm of 1,000 Fable 5.2 agents completed the step-by-step translation of Fermat’s Last Theorem in 30 hours. This was a major milestone, as the Opus 5.5 model reached only 26% completion before stopping.

Robotics and biosecurity

Beyond software and math, Fable 5.2 showed strong zero-shot physical dexterity in the RoboDojo simulation benchmarks. It scored 64.9% in controlling simulated two-arm robotic arms without needing task-specific pre-training. This suggests a high level of generalizable reasoning for physical movements.

Anthropic is also using these models for complex biological modeling. Mythos 5.2 was able to accurately model 28 of 34 complex gene-therapy protein capsids. This level of precision is critical for biomedical research and the development of new therapies.

To manage risks, Anthropic has created a split between public and private access for these models. Fable 5.2 includes public safeguards against cyber and biological misuse. However, Mythos 5.2 allows vetted institutions to access identical core weights under specific verification programs.

Strategic market positioning

As reported by PaidFreeDroid, Anthropic is using a deliberate strategy to balance capability with safety. By offering Fable 5.2 as a specialized reasoning engine, they are targeting industrial users who need high-effort outputs. This distinguishes it from the more general-purpose models used by everyday consumers.

The move to lower API prices while maintaining high performance creates a competitive environment for other frontier models. It allows companies to build complex agentic workflows that require high-level reasoning without prohibitive costs. This strategy focuses on making the models useful for long-horizon autonomous tasks.

Looking ahead, the success of these models will likely depend on their adoption in hardware and life sciences. The ability to model protein capsids and perform hardware CAD modeling opens new doors for industrial research. Anthropic is positioning itself as a leader in these high-stakes, technical domains.