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 currently 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, they employ different safety measures for different use cases.

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

In other words, one is designed for real-world applications, while the other remains in a more strictly controlled environment. This may also be Anthropic's exploration of the future shape of AI products: the most powerful models do not necessarily have to be offered in the same form to everyone.

The official statement indicated that these two models represent a significant advancement in the capabilities of the Claude series and demonstrate how they continue to push for safe deployments while enhancing model capabilities.

As the "strongest model on the planet," Fable 5.1 continues to perform impressively; official data shows that Fable 5.1 outperforms its predecessor, flagship model Fable 5, in various benchmark tests.

Stronger capabilities come with a lower price. Anthropic stated that Fable 5.1 maintains the same base price, while the cost of cache reading has been reduced by 75% compared to Fable 5. In practical use, this lowers the overall cost of the model for typical workloads by about 25%; for highly agent-oriented workloads, costs can be reduced by up to 45%.

It seems that even the strongest models are emphasizing cost-effectiveness.

Let's take a closer look.

Lower tier catches 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 accomplish tasks that previously required the high tiers of Fable 5 at a lower inference cost.

According to official testing, Fable 5.1 can achieve performance close to or even exceeding that of Fable 5 at low to moderate inference intensities. Claude Code defaults to high inference intensity, while Claude Cowork and Claude .ai default to moderate intensity. Users can also adjust according to the complexity of the task, balancing speed, cost, and effectiveness.

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

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

In multiple benchmark tests, Fable 5.1's performance surpassed that of Fable 5, with the most noticeable improvements in scientific research-related agents and business workflow capabilities.

In the Terminal - Bench 4.0 test, Fable 5.1 and Mythos 5.1 both 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 an important change in Fable 5.1 is its greater willingness to trace the root causes of issues, making it better suited for executing complex tasks that last hours or even tens of hours.

Investment firm Millennium found during testing that a piece of internal code crashes approximately once every million executions. This issue troubled the engineering team for four to five years, and other models were unable to identify the cause. Fable 5.1 located the issue by disassembling external vendors' libraries and comparing core dump files, ultimately pinpointing a bug in a third-party library.

Ramp also had Fable 5.1 run continuously for 38 hours. After discovering that its original machine learning results were affected by label errors, the model self-corrected the problem, concurrently launching six sets of experiments, ultimately organizing new results and subsequent recommendations.

From writing code to conducting research

In this release, Anthropic devoted a considerable amount of space to discuss the performance of Fable 5.1 and Mythos 5.1 in scientific research.

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

For three target proteins, the binder affinity designed by Mythos 5.1 reached ten times that of the best solution from the Adaptyv Bio-related protein design competition. When faced with 12 target proteins, the model's effective binder hit rate was nearly 50%, while the industry standard provided by Anthropic is 10% to 15%.

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

NASA's Magellan spacecraft captured radar images of Venus, with a small shield volcano about 15 kilometers in diameter in the center.

The original maps could only display details at scales of 10 to 20 kilometers, while the new map improves resolution to 2 to 3 kilometers, with elevation measurement accuracy increased by up to 25%. Anthropic plans to make this map publicly available under a knowledge-sharing license for reference in subsequent missions like NASA 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 genome deep learning models by up to 2.5 times through 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 genome models, Mythos 5.1 increased inference speed 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 drop by 75%, agent tasks can be up to 45% cheaper

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

The input and output prices for Fable 5.1 remain unchanged, still set at $10 and $50 per million tokens, but the cache reading price has been reduced by 75% from previous levels to $0.25 per million tokens.

Cache reading refers to the model's reuse of already processed context. Its proportion may be limited in general Q&A, but in agent tasks needing repeated access to code repositories, historical dialogues, and tool results, caching often constitutes the main cost.

According to Anthropic’s estimates based on actual usage data from August 2026, the overall cost of running typical tasks with Fable 5.1 is about 25% lower than with Fable 5; for complex agent tasks with longer contexts and frequent tool calls, cost reductions can be as high as about 45%.

With the cache reading price reduced, the cost for typical tasks with Fable 5.1 drops by approximately 25%, while costs for highly agent-oriented tasks decrease by up to about 45%.

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

Strengthened anti-distillation mechanisms

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

Fable 5.1 is fully opened to general users and enterprises; Mythos 5.1 has more lenient restrictions for cybersecurity and life sciences, currently only provided to vetted individuals and institutions.

In the cybersecurity domain, Fable 5.1 can already help users identify 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 cybersecurity protective mechanism has reduced false positives by about 60% compared to Fable 5. The biosafety mechanism has also lowered the false positive rate for fundamental biology and medical issues by 85%, with advanced capabilities related to life sciences research mainly 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 companies responsible for default manual reviews, achieving near "zero data retention" privacy while retaining security monitoring capabilities. This mechanism is planned to be implemented in phases starting this fall.

For large-scale model distillation, Fable 5.1 has also implemented new restrictions. API accounts created after that day will not be able to manually modify previous contexts in multi-turn dialogues while retaining Claude's historical thinking records. Anthropic states that this change is primarily aimed at blocking a publicly known method of capability extraction.

Additionally, to comply with the EU AI Act requirements, Fable 5.1's text output will contain invisible watermarks. Anthropic has also begun a small-scale test for detection APIs, initially open to regulatory agencies, media, fact-checking organizations, and research institutions.

Finally, as the netizens said, the early bird gets to use 5.1 first.