Anthropic releases "Claude Fable 5.2" and "Claude Mythos 5.2," cuts API prices to $8 input and $40 output; tops overall index in Artificial Analysis evaluation

Two models for the hardest technical work updated

Anthropic released the AI models "Claude Fable 5.2" and "Claude Mythos 5.2" in an official blog post dated Nov. 9. The company described them as models "designed to tackle the most daunting technical challenges" and said they improve on coding, reasoning and image understanding. The update follows "Fable 5.1" and "Mythos 5.1," released on Sept. 1.

Fable and Mythos share the same underlying model. Fable has strong safeguards to prevent misuse in cybersecurity and biology, and Anthropic said these safeguards allow it to be offered to a wide range of users. Organizations working on beneficial uses in cybersecurity or biomedical research apply to use Mythos through the "Life Sciences Verification Program" and the "Cyber Verification Program."

According to testers, 5.2 explains things more clearly than earlier Fable and Mythos models and generates output up to 30% faster than Fable 5.1 and Mythos 5.1. Anthropic said Fable and Mythos are suited to ambitious technical work without set answers and recommended starting most coding and knowledge work with "Claude Opus 5.5." Depending on the scope and complexity of the work, users can move up to Fable or Mythos, or down to "Claude Sonnet 5.5" or "Claude Haiku 5.5."

Claude's official account announced the two models on X at 3:57 a.m. Nov. 10 Japan time.

In early November, a series of posts on X showed animations said to have been made with "Fable 5.5", and an image purporting to show the name "Mythos 5.5" in the source of the Mythos product page also circulated. The version number Anthropic released this time was 5.2.

API pricing cut to $8 input and $40 output

API pricing is $8 per million input tokens and $40 per million output tokens, down from $10 input and $50 output for Fable 5.1 and Mythos 5.1. Reading cached input costs $0.20. According to the developer documentation, cache writes cost $10 for a 5-minute retention and $16 for a 1-hour retention, and batch processing costs $4 input and $20 output.

With the optional "fast mode," token generation is up to 2.5 times faster, priced at $16 input and $80 output. A setting that processes data only within the United States costs 1.1 times the standard price for both input and output. In the model list in the developer documentation, Opus 5.5, which Anthropic recommends using first, is priced at $4 input and $20 output.

49.1% on FrontierCode 2.0, which Anthropic calls the highest among released models

In the evaluation table Anthropic published, Fable 5.2 scored 49.1% with reasoning depth set to "Xhigh" on FrontierCode 2.0, which measures whether a model can produce code changes that open-source developers actually accept (merge). In the same table, Opus 5.5 scored 45.5%, OpenAI's "GPT-6 Astra" 42.8% and Google's "Gemini 4 Carbon" 41.1% (with reasoning at Medium). Unless noted, the figures in the table are for the deepest reasoning setting available for each model, and Fable 5.2, Fable 5.1 and Opus 5.5 results are with safeguards enabled unless noted.

Anthropic said Fable 5.2's score with reasoning at "High" was almost the same as Opus 5.5's best score (45.3% and 45.5%), at about half the cost per task. The 49.1% at Xhigh is the highest among the released models the company tested, it said.

In the same table, however, Gemini 4 Carbon's 53.9% beat Fable 5.2's 51.4% on AutomationBench, which automates business workflows. On AA-Briefcase v1.1, which measures knowledge work, Fable 5.2's rubric-based score of 66.1% exceeded Opus 5.5's 61.2%, but its head-to-head rating (Elo) was 1785 against Opus 5.5's 1800.

Anthropic also showed examples of use from pre-release testing. About 1,000 Fable 5.2 agents completed in 30 hours the "formalization" of the proof of Fermat's Last Theorem, rewriting it in a form a computer can verify step by step, the company said. The company's first formalization, by an unreleased research version of Claude, had taken 11 days with a dedicated setup and guidance from researchers.

The company revised this part of the official blog on Nov. 11 and added a footnote. According to the footnote, the agents were given a list of 60 propositions, a one-sentence outline of the classical proof strategy, the definitions used in the propositions and 47 lemmas already proved in the earlier formalization, and were assigned lead and supporting roles. It said no other instructions were given on how to proceed with the proof or how to organize.

In hardware design, the model was given a one-paragraph instruction and designed three electronic devices — a relay unit for a home voice assistant, a USB speakerphone and a handheld virtual pet — including circuit boards, 3D-printed enclosures and software. The boards were manufactured in cooperation with Diode, which designs and manufactures circuit boards, and all three worked without electrical rework, the company said.

Tops the overall index in AA's evaluation, but deepest setting costs about 2.6 times Opus 5.5

Artificial Analysis (AA), a U.S. firm that independently evaluates AI models, published its results on X at 4:45 a.m. Nov. 10 Japan time. In AA's evaluation, Fable 5.2 ranked first on the overall "Intelligence Index v5." At the deepest reasoning setting (max), it scored 65, ahead of Gemini 4 Carbon (high) at 60, Claude Opus 5.5 (max) at 59 and GPT-6 Astra (max) at 57. The previous Fable 5.1 (max) scored 49.

According to AA, Fable 5.2 ranked first on five of the nine evaluations that make up the index. Even with reasoning set to high and xhigh, it scored 62, higher than any other model at any setting. At the high setting, it used about 40% fewer output tokens than Opus 5.5 (max), AA said.

However, the cost per task in AA's evaluation was $26.5 at max, about 2.6 times the $10.0 for Opus 5.5 (max) and about twice the $12.99 for Fable 5.1 (max). AA said gains were small in agentic knowledge work and terminal-based coding. On AA-Briefcase, it scored 1785, within the confidence interval of Opus 5.5's 1800 and slightly below Claude Sonnet 5.5's 1813. On Terminal-Bench 4.0, it scored 60%, tying Opus 5.5 and trailing Moonshot AI's "Kimi K3.1 Pro" at 67%, Gemini 4 Carbon at 65% and Sonnet 5.5 at 64%.

Fable offered on 5 paid plans and the API; Mythos for organizations that pass a review

According to Anthropic's product page, Fable 5.2 is available to users on the Pro, Max, Ultra, Team and Enterprise plans. For developers, it is offered on the Claude Platform and various marketplaces, as well as Amazon Web Services (AWS), Google Cloud and Microsoft Foundry. The official blog says it is available from the same day on all platforms, including AWS, Google Cloud and Microsoft Azure.

The API model name is claude-fable-5-2. According to the developer documentation, on Amazon Bedrock it is global.anthropic.claude-fable-5-2 or us.anthropic.claude-fable-5-2, available only through the bedrock-runtime endpoint. The context window is 1 million tokens, and a single output can be up to 128,000 tokens (128K). Reasoning depth (effort) has five levels, with high as the API default, and "adaptive thinking," in which the model adjusts how much it thinks, is always on. 30-day data retention is required for safety monitoring.

Mythos 5.2 is available to organizations that pass a review in the verification programs, with pricing starting at $8 input and $40 output. The API model name is claude-mythos-5-2, and on Amazon Bedrock it is global.anthropic.claude-mythos-5-2 or us.anthropic.claude-mythos-5-2. Claude Enterprise customers can use Mythos 5.2 through "Claude Security," which searches for vulnerabilities in code, even if they are not in a verification program.

In the week of release, Anthropic is adding extra Fable usage for subscribers on eligible paid plans. Plans and seats that include Fable get one to three resets of Fable usage, depending on the contract period. Standard seats on the Pro and Team plans can use Fable for up to 50% of the weekly usage limit.

Anthropic's official account also posted on X at 4:00 a.m. Nov. 10 Japan time that the two models are now available.