Anthropic Releases Claude Fable 5.2 and Mythos 5.2, Cutting API Prices by 20%

On November 9, 2026, Anthropic announced the AI models Claude Fable 5.2 and Claude Mythos 5.2. The company says it has improved coding and image understanding, and enhanced performance on complex reasoning and long-running agentic work. Compared with Fable 5.1 and Mythos 5.1 respectively, output generation is up to 30% faster, and the per-token prices for both input and output on the standard API have dropped by 20%.

Introducing Claude Fable 5.2 and Claude Mythos 5.2 - Anthropic

Fable 5.2 is available on all of Claude's paid plans, but usage limits and whether extra charges apply differ by plan. In addition to the Claude API, it is offered on AWS, Google Cloud, and Microsoft Foundry. Mythos 5.2 is provided to organizations and individuals working in cybersecurity and biomedical research who have passed review under the company's verification program.

Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.

They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision. https://t.co/jP0AutLxXb— Claude (@claudeai) November 9, 2026

Publishes evaluation results for coding, reasoning, and more

Anthropic published initial evaluations on FrontierCode 2.0, which measures whether OSS maintainers accept generated code changes. Fable 5.2, with effort—which adjusts the amount of compute spent on reasoning—set to High, scored 45.3%. The company says it achieved a result close to the 45.5% of Opus 5.5 at Max, at about half the cost per task. Setting Fable 5.2 to Xhigh raised the score to 49.1%. All evaluations were run with safety measures enabled.

The company also published evaluation results for scientific research, computer operation, reasoning, and more. Including the results of Terminal-Bench 4.0 from the external evaluation organization Vals AI, the comparison with the previous generation Fable 5.1, Opus 5.5, and OpenAI's GPT-6 Astra is as follows.

Benchmark[1] Fable 5.2 Fable 5.1 Opus 5.5 GPT-6 Astra

FrontierCode 2.0 49.1% 32.4% 45.5% 42.8%

FrontierSWE v2 69.9% 56.3% 62.3% 65.5%

Terminal-Bench 4.0[2] 67.6% 55.8% 66.4% 57.9%

Terminal-Bench-Science 0.1 74.6% 52.6% 58.7% 68.1%

AutomationBench 51.4% 31.4% 42.5% 41.4%

OSWorld 2.1 89.3% 83.5% 87.2% 79.5%

Humanity’s Last Exam 74.2% 65.6% 67.7% 57.2%

Chartography 73.4% 44.8% 64.4% 71.0%

Publishes evaluation results for Mythos 5.2, also integrated into Claude Security

Fable and Mythos share the same underlying model, but their safety measures for cybersecurity and biology differ, and evaluation results differ on tasks affected by those measures. The published figures for Mythos 5.2, together with the corresponding results for Fable 5.2, are summarized in the following table. For CyScenarioBench, the results for OpenAI's published GPT-6 Cyber are also listed.

Benchmark[3] Fable 5.2 Mythos 5.2 GPT-6 Cyber

Terminal-Bench 4.0 (Claude Code) 67.6% 70.5% —

Terminal-Bench-Science 0.1 74.6% 75.3% —

CyScenarioBench — 76.3% 61%

CVE Rediscovery — 89.6% —

De novo binder design (24 hours) — 85.1% —

Anthropic also reported improvements in Mythos 5.2's ability to discover, reproduce, and fix vulnerabilities. CVE Rediscovery in the table is an evaluation conducted in October 2026 on 1,829 CVEs, in which it found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1.

Mythos 5.2 has also been added to Claude Security, which examines code for vulnerabilities. It is available in a public beta for Claude Enterprise, and participation in the Cyber Verification Program is not required. Prior to this, on October 21, the company also announced feature updates for Claude Security, including continuous scanning.

Claude Security now scans continuously and tracks fixes - Claude

This update added the ability to continuously examine repositories and re-check for vulnerabilities as code changes. When a fix is reflected in the code, the status of detected vulnerabilities is updated to resolved, and if they recur, it reverts to unresolved. It also reflects team-registered operating environments and security policies in assessments, and can integrate with the development tools teams normally use via API or MCP.

Applied to electronic device design and scientific research

At Anthropic's own development site, an agent team using an early version of Fable 5.2 migrated a service actually in operation from Python to Rust, and the service reportedly started about 26 times faster.

During pre-release testing, Claude Fable 5.2 and Claude Mythos 5.2 worked on ambitious projects in many different fields. Here are a few.

From a one-paragraph prompt, Fable 5.2 designed three working electronic devices, including this home voice assistant hub. https://t.co/efOYdhRr7K— Claude (@claudeai) November 9, 2026

Outside software development, the company also introduced an experiment in which Fable 5.2 was given a short prompt and asked to design a home voice assistant hub, a USB speakerphone, and a portable virtual pet. The model went from drafting specifications to selecting components, designing the routing of a four-layer circuit board, designing the enclosure to be made with a 3D printer, and implementing the software.

When the partner company Diode manufactured and assembled these designs, all three models reportedly worked without any circuit modifications.

In the biomedical field, Mythos 5.2 devised a method to predict the structure of the entire capsid—the shell of the adeno-associated virus (AAV) used in gene therapy. It wrote code that has existing open-source models predict the structure of each part and combines the results to model the entire shell. When Anthropic tested it with two types of prediction models, it was able to accurately model 28 of the 34 targets[4].

In chemistry, Claude also introduced on its official X an experiment in which it proposed synthetic routes for organic molecules. External collaborating researchers rated Claude's proposals as equivalent to the level of a student who has just entered graduate school.

Formalizes mathematical proofs, also achieves results on unsolved problems in theoretical physics

In mathematics, 1,000 agents using an early version of Fable 5.2 reportedly formalized the proof of Fermat's Last Theorem in about 30 hours. This is an effort to rewrite a proof already shown by mathematicians into a form in which a computer can verify each step of the logic, using Lean, a language for writing and verifying proofs. Based on propositions and proof strategies prepared by humans, the agents divided the work among themselves. The detailed division of labor is described in the column at the end.

In theoretical physics, Anthropic reported that Fable 5.2 solved two unsolved problems concerning black holes and quantum gravity.

One concerns calculations related to the "Hawking-Page transition," in which a black hole forms when the temperature is raised. The challenge was whether, in a specific theoretical model, this change occurs abruptly even when quantum effects are taken into account. Claude advanced a calculation that had been unfinished since 2003 and reportedly obtained a result showing that even under conditions of weak interaction, a "first-order phase transition"—an abrupt change—occurs. The details of the calculation have been compiled in a published paper.

The other is research that determined the limits of a computational method for narrowing down candidates that could hold as theories of quantum gravity. Until now, only a range was known for the numerical value representing the strength of the constraints obtained by this method. Claude reportedly derived that value exactly and clarified a limit that cannot be exceeded even if the method is improved. A paper on this has also been published [5].

Everyday work with Opus, large-scale migrations and research with Fable

Anthropic recommends that for many coding and knowledge tasks, users start with Opus 5.5 and choose a model according to the difficulty of the task, cost, and response speed. It cites large-scale refactoring and migrations, architecture considerations, supervision of large numbers of sub-agents, and scientific research and quantitative analysis as the main uses of Fable 5.2.

Which Claude model should you build with? Fable, Opus, Sonnet, or Haiku - claude.dev Blog

In its selection guide, it lists Sonnet 5.5 as a candidate for fixes and repetitive work with a clearly defined scope of changes, and Haiku 5.5 for large-scale classification and extraction. It also presents a configuration in which multiple models are combined, with tasks such as file searches and reading logs delegated to sub-agents using low-cost models, while Fable 5.2 manages the whole. It recommends trying to handle things by adjusting effort before switching models.

Anthropic's Felix Rieseberg also cited, in a post on X, rewriting an entire codebase, drafting the first version of a large deliverable, and examining problems whose requirements are not yet clear as examples of what Fable excels at.

API unit price cut by 20%; the cost of an entire task also depends on usage

The standard API pricing for Fable 5.2 and Mythos 5.2 is $8 per million tokens for input and $40 for output. Both were lowered by 20% from 5.1's $10 input and $50 output.

What's new in Claude Fable 5.2 - Claude Platform Docs

However, Anthropic's Addy Osmani added in a post on X that Fable 5.2 uses many tokens, and that in their tests the cost per task tended to be roughly the same as with 5.1.

A Fast mode is also available that makes output token generation up to 2.5 times faster than normal. Pricing is $16 per million tokens for input and $80 for output, and it is available in Claude Code, Claude Platform, supported services on Google Cloud, and Microsoft Foundry hosted by Anthropic. However, Mythos's Fast mode can only be used by organizations participating in the company's verification program.

Looking at the basic specifications of both models, the context window is 1 million tokens, and they can output up to 128K tokens. According to section 1.1.1 of the system cards for both models, the knowledge cutoff is June 2026.

In the API, Adaptive thinking, which reasons according to the task, is always enabled. The default effort value is set to High in the API and Medium in the Claude app and Claude Code.

Developers of existing systems that build API requests themselves need to check how conversation history is sent during migration. If the system prompt or past messages are changed, the inference data from the point of change onward becomes invalid. If this data is left in the request and resent, the API's default behavior will result in an error. Also, Fable 5.2 and Mythos 5.2 can read each other's inference data, but if you switch to any other model, inference data generated by 5.2 is not carried over. Specific changes can be confirmed in the migration guide.

Prompt guide for 5.2 released

The company has also released a prompt guide tailored to the characteristics of Fable 5.2 and Mythos 5.2. It explains how to adjust instructions and the execution environment for issues such as long-running work ending partway through, being asked for approval more than necessary, progress not being displayed, and overuse of subagents.

Prompting Claude Fable 5.2 - Claude Platform Docs

The guide recommends that developers using the API, when migrating, first try setting effort to High, and also check quality and cost with Medium and Xhigh. On the other hand, it says that instructions such as "answer directly without thinking" had no effect in suppressing reasoning when effort was Medium or higher. As a way to reduce reasoning, it introduces the setting of lowering effort.

When carrying over system prompts, skills, and notes intended for previous models, it recommends trimming cautions that are no longer needed and clarifying which operations are prohibited and which require approval. At that time, it cautions that rewriting past instructions mid-conversation can cause an error in the API, so users should try it in a new conversation.

In addition, it includes example instructions that encourage long-running work to continue to the end. When the company added this instruction, set effort to Medium, and re-ran recorded sessions of a coding agent, the number of sessions that ended even though there was still work that could be done was roughly halved, while output tokens increased by about 30%. To check such effects and changes in consumption, the guide recommends trying one instruction change at a time. It also advises that, in such cases, a mechanism to confirm with the user be retained for dangerous or irreversible operations.

Fable's safety measures in the cybersecurity and biology fields block most requests for high-risk cyber-related work such as penetration testing, as well as most requests related to biological research. When using Mythos in these fields, apply through the Cyber Verification Program or the Life Sciences Verification Program. Among these, Defense Access in the cybersecurity field allows individuals such as independent researchers and OSS maintainers to apply.

The alternative models for requests blocked by Fable are Opus 4.8 for cybersecurity and Opus 5 for biology. To use alternative responses in the API, it is necessary to configure or implement retries to a different model.

Meanwhile, improvements have also been made to reduce unnecessary blocks. In initial tests on claude.ai, Claude Code, and Cowork, the number of blocks decreased by about 55% compared with Fable 5.1.

Also, in training both models, the company intentionally suppressed improvements in some research capabilities that could directly accelerate AI development. As a result, for things like optimizing processing for accelerators and improving the structure of large language models, performance is said to be about the same as Opus 5.5. It says capabilities for ordinary coding and hardware development, etc., have not been affected.

In safety evaluations, Mythos 5.2 reportedly exhibited behaviors such as reporting completion when work was not finished and omitting important information less often than Opus 5.5, Sonnet 5.5, and Mythos 5.1.

Using both models also comes with data retention conditions for safety monitoring. Anthropic generally retains prompts and outputs for 30 days, and use under Zero Data Retention (ZDR), which does not retain data, requires the company's explicit approval. On the other hand, Enterprise Frontier Safeguards (EFS), offered to enterprises in private beta, stores prompts and outputs on the customer's infrastructure for safety monitoring, allowing operation without Anthropic retaining data. In this announcement, the company informed eligible customers that they can continue using ZDR even before adopting EFS.

Release benefits include Fable trial quota and usage limit resets

As a release benefit, on Pro and Team Standard seats, all Fable models can be used at no additional charge, up to 50% of the weekly usage limit, until 16:59 Japan time on November 17, 2026. After reaching the limit or after the period ends, if the user has enabled usage credits for pay-as-you-go billing, Fable is used with those credits. The default model does not change, and users select Fable from the model selection menu.

Claude Fable 5.2 launch offers - Claude Help Center

On the other hand, on Max, Ultra, Team Premium seats, and Premium seats of the old Enterprise plan billed per seat, Fable usage is included in the regular plan. The weekly usage limit is shared with other models, and up to 50% of it can be used for Fable. Consumption is faster than with other models.

In addition, for eligible existing subscribers, a Fable limit reset is provided that restores the portion of the weekly usage limit used by Fable. The usage limit itself does not increase. For accounts with a 5-hour usage limit, that limit is also restored, including usage of other models.

Pro and Team Standard seats can use it once, with a deadline of 16:59 Japan time on November 17. Max, Ultra, Team Premium seats, and Premium seats of the old Enterprise plan billed per seat can use it 1 to 3 times depending on the contract period, with a deadline of 16:59 Japan time on March 9, 2027 [6].

Users can check the number of resets available to them and their expiration dates in the "Usage" section of Claude's settings screen.