Just now, Claude 5.1 has been released! The world's strongest model is here

Original Follow AI's Machine Heart
Just now, Anthropic launched the next generation models Claude Fable 5.1 and Claude Mythos 5.1.

Anthropic claims that Claude Fable 5.1 is one of the most powerful publicly available models, designed for complex reasoning, long-duration tasks, and agent workflows. Claude Mythos 5.1 represents their exploration of higher capability boundaries.

It is important to note that while both share the same foundational model capabilities, different safety measures have been adopted for different usage scenarios.

Fable 5.1 is open to general users, developers, and enterprises, while Mythos 5.1 is aimed at vetted high-risk domain partners, maintaining stricter access controls.

In other words, one is responsible for entering the real-world tasks, while the other is kept in a more strictly controlled environment. This may also be Anthropic's exploration of the future forms of AI products: the strongest models may not necessarily be provided to everyone in the same form.

The official statement indicates that these two models represent significant advancements in the capabilities of the Claude series and demonstrate how they continue to advance safe deployment while enhancing model capabilities.

As the "strongest on the surface" model, Fable 5.1 still performs impressively. Official data shows that Fable 5.1 outperforms the previous flagship Fable 5 in multiple benchmark tests.

However, with greater capabilities, the price has decreased. Anthropic states that Fable 5.1 maintains the same base price, while the cost of cache reads has decreased by 75% compared to Fable 5. In practical use, this reduces the overall cost of the model for typical workloads by about 25%; for highly agentized workloads, costs can be reduced by up to 45%.

It seems that even the most powerful models are focusing on cost-effectiveness.

Let's take a closer look.

Lower tiers catch up to the previous generation, Fable 5.1 focuses on "cost-effectiveness."

Rather than simply refreshing the leaderboard, Anthropic wants to emphasize this time: Fable 5.1 can complete tasks that previously required the high tiers of Fable 5 with lower inference costs.

According to official tests, Fable 5.1 has already achieved performance close to or even exceeding that of Fable 5 at low to medium inference intensities. Claude Code defaults to high inference intensity, while Claude Cowork and Claude .ai default to medium intensity, allowing users to adjust between speed, cost, and effectiveness based on task complexity.

Specifically, Fable 5.1 achieved a score of 52.6% in the scientific research agent benchmark Terminal - Bench - Science 0.1, more than doubling Fable 5's 24.7%; in Terminal - Bench 4.0, Fable 5.1 reached 55.8%, with the more capable Mythos 5.1 further achieving 60.9%.

In knowledge work, computer operations, and business process tests, the new model also showed improvements. AutomationBench scores increased from 17.1% for Fable 5 to 31.4%, and CursorBench 3.2.0 improved from 70.5% to 73.4%.

In multiple benchmark tests, Fable 5.1's scores exceeded those of Fable 5, with the most significant improvements seen in scientific research agents and business workflow capabilities.

In the Terminal - Bench 4.0 test, both Fable 5.1 and Mythos 5.1 surpassed the previous generation Mythos 5 at different inference intensities.

In the Terminal - Bench - Science 0.1 test, Fable 5.1 achieved a maximum score of 52.6%, more than double that of Fable 5.

Anthropic believes that an important change in Fable 5.1 is its greater willingness to track the root causes of problems, making it more suitable for executing complex tasks that last for hours or even tens of hours.

Investment firm Millennium found in testing that a piece of internal code crashes approximately once every million runs. This issue had troubled the engineering team for four to five years, and other models had not identified the cause. Fable 5.1 pinpointed the problem to a bug in a third-party program library by disassembling the external vendor's library and comparing core dump files.

Ramp also had Fable 5.1 run continuously for 38 hours. After discovering that the original machine learning results were affected by labeling errors, the model autonomously corrected the issue and concurrently initiated six sets of experiments, ultimately compiling new results and subsequent recommendations.

From writing code to conducting scientific research

In this release, Anthropic devoted a significant portion to discussing Fable 5.1 and Mythos 5.1's performance in scientific research.

In protein design experiments, researchers had Mythos 5.1 call upon open-source protein design and folding tools, then sent the generated designs to two external institutions for experimental validation.

For three target proteins, the binding affinity of the ligands designed by Mythos 5.1 reached ten times that of the best solutions in the Adaptyv Bio-related protein design competition. When faced with twelve target proteins, the model's effective ligand hit rate approached 50%, while the industry common level provided by Anthropic is 10% to 15%.

Fable 5.1 also utilized radar images collected by NASA's Magellan spacecraft over 30 years ago to generate a new high-resolution elevation map for about one-third of Venus's surface.

Radar images of Venus taken by NASA's Magellan spacecraft, with a small shield volcano approximately 15 kilometers in diameter at the center.

The original map could only present details at a scale of 10 to 20 kilometers, while the new map improves resolution to 2 to 3 kilometers, with height measurement accuracy improved by up to 25%. Anthropic plans to publicly release this map under a knowledge-sharing license for future missions such as NASA's VERITAS and the European Space Agency's EnVision.

In the field of computational biology, Mythos 5.1 improved the running speed of seven open-source protein and genomic deep learning models by up to 2.5 times by writing custom GPU kernels and caching intermediate results, while maintaining consistent output results. In some whole-genome analysis tasks, it is expected to reduce GPU costs by 30% to 60%.

After optimizing seven open-source protein and genomic models, Mythos 5.1's inference speed increased by 1.4 to 2.5 times.

Anthropic aims to use these cases to demonstrate that the model is evolving from organizing information and writing code to proposing solutions, running tools, and delivering research results that can be experimentally validated.

Cache prices dropped by 75%, with agent tasks becoming up to 45% cheaper.

Beyond performance, price is the most direct change in Fable 5.1.

The input and output prices of Fable 5.1 have not been adjusted, remaining at $10 per million tokens and $50, respectively, but the cache read price has been reduced by 75%, down to $0.25 per million tokens.

Cache reading refers to the model reusing already processed context. Its proportion may be limited in ordinary Q&A, but in agent tasks that require repeated reading of codebases, historical dialogues, and tool results, caching often constitutes a major cost.

According to Anthropic's calculations based on actual usage data from August 2026, the overall cost of running typical tasks with Fable 5.1 is about 25% lower than that of Fable 5; for complex agent tasks with longer contexts and frequent tool calls, the cost reduction can reach up to about 45%.

After the reduction in cache read prices, the cost of typical tasks with Fable 5.1 is reduced by about 25%, while the cost of highly agentized tasks is reduced by up to about 45%.

Fable 5.1 is now available on Claude .ai, Claude Code, and Claude API, and is also provided through AWS, Google Cloud, and Microsoft Azure. Developers can call the new model via claude - fable -5-1.

Strengthened anti-distillation mechanism

Fable 5.1 and Mythos 5.1 actually use the same underlying model, with the main difference being safety restrictions.

Fable 5.1 is fully open to general users and enterprises; Mythos 5.1 has more relaxed cybersecurity and life sciences restrictions and is currently only available to vetted individuals and institutions.

In the field of cybersecurity, Fable 5.1 can already help users discover software vulnerabilities but still cannot directly generate exploit programs. Dual-use tasks such as penetration testing, exploit generation, and binary file-based vulnerability scanning may still be referred to models with stricter restrictions.

After adjustments, the new version of the cybersecurity protection mechanism has reduced false positives by about 60% compared to Fable 5. The false positive rate for biosafety mechanisms on basic biology and medical issues has also decreased by 85%, while advanced capabilities related to life sciences research are primarily opened through the Mythos 5.1 review program.

Anthropic has also launched Enterprise Frontier Safeguards. Enterprises can store data in their controlled cloud infrastructure, with the enterprise responsible for default manual reviews, thus achieving near "zero data retention" privacy effects while retaining security monitoring capabilities. This mechanism is planned to be phased in starting this fall.

For large-scale model distillation, Fable 5.1 has also added new restrictions. API accounts created after that day will not be able to manually modify previous context in multi-turn dialogues while retaining Claude's historical thought records. Anthropic states that this change is mainly to block a publicly known capability extraction method.

Additionally, to meet the requirements of the EU's Artificial Intelligence Act, Fable 5.1's text output will include invisible watermarks. Anthropic has also begun small-scale testing of detection APIs, initially open to regulatory agencies, media, fact-checking organizations, and research institutions.

Finally, netizens are right; the early bird gets to use 5.1 first.