Anthropic Releases Claude Fable 5.2 with 20% API Price Cut

宝玉 (@dotey) · 2026-11-09 19

Claude Fable 5.2 Released: One Tier Stronger Than Opus 5.5, API Price Cut by 20%; Max and Ultra Get 1 Fable Reset, Pro Can Also Use Fable Directly This Week

On November 9, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, its highest-end model tier. According to the official announcement, the new model is one tier stronger than Opus 5.5 on nearly all tested capabilities; output speed is up to 30% faster than 5.1; and API pricing drops from $10 /$50 per million input / output tokens to $8 /$40.

For subscribers, usage on Max ($100/month), Ultra ($200/month), and Team premium seats is unchanged: Fable 5.2 is included in the plan just like the previous Fable, with up to half of the weekly quota usable on Fable, and this time an additional 1 to 3 "Fable quota resets" are included.

For Pro ($20/month) and Team standard seats, Fable normally does not draw from plan quota and requires enabling "extra usage" for pay-as-you-go billing. From November 9 to 16, this temporarily changes for one week so that usage is deducted from plan quota at no extra charge, capped at half the weekly quota, plus 1 quota reset; pay-as-you-go billing resumes on the 17th. Free users cannot use it.

[What Is the Relationship Between Fable and Mythos]

The two are the same model, differing only in guardrails.

Fable adds cybersecurity and biology guardrails, which is why it can be opened to all paid users. Penetration testing, writing exploit code, analyzing binary vulnerabilities, and most biology research tasks will be blocked and answered instead by Opus 4.8 (cybersecurity) or Opus 5 (biology). According to the official announcement, Fable 5.2 has about 55% fewer false blocks than 5.1.

Mythos does not have this layer of restriction and is only available to vetted organizations. Organizations doing cybersecurity or biopharmaceutical research can apply through Anthropic's "Cybersecurity Verification Program" and "Life Sciences Verification Program." What ordinary users can access is Fable.

[How to Use Quota Resets, and What Else to Watch Out For]

A quota reset refunds the quota already spent on Fable this week; quota spent on other models is not refunded; for accounts with a 5-hour limit, the entire 5-hour window is refilled. It does not raise the cap, and Fable still can account for at most half of the weekly quota.

How many resets Max, Ultra, and Team premium seats receive depends on subscription length: 1 for less than six months, 2 for six months to one year, and 3 for more than one year, to be used by March 8, 2027. The 1 reset for Pro and Team standard seats must be used by November 16. Only accounts that had already paid before 10 a.m. Pacific Time on November 9 are eligible; click "Free Reset" in "Settings > Usage" to use it.

Fable consumes quota faster than other models. The default model will not switch automatically; you need to manually select Fable 5.2 in the model menu. In the Claude app and Claude Code, the default thinking effort is Medium. Claude Code needs to be upgraded to version 2.1.321 or higher. In Enterprise, Fable 5.2 is initially off by default and must be enabled by an administrator; after 7 days it will be automatically added to roles that already have Fable enabled, and administrators can choose in advance not to add it.

Developers use the model name claude-fable-5-2, and it is already available on AWS, Google Cloud, and Microsoft Azure. There is also an optional fast mode, with generation speed up to 2.5 times faster, at double the price of $16 /$80 per million tokens.

[Is It Worth Switching]

Anthropic's own recommendation is: for most coding and knowledge work, start with Opus 5.5, switch to Fable or Mythos when tasks are larger and more open-ended, and downgrade to Sonnet 5.5 or Haiku 5.5 for simple tasks. Subscribers also need to account for quota: Fable consumes quota faster, and at most half of the weekly quota can be used on it.

Per token, Fable 5.2 still costs twice as much as Opus 5.5 ($4 /$20). But it uses fewer tokens to complete the same task. On the agentic coding benchmark FrontierCode 2.0, Fable 5.2 at High effort scores 45.3%, essentially on par with Opus 5.5 at the highest Max effort at 45.5%, while costing only about half as much per task. At Xhigh effort it scores 49.1%, the highest among released models; GPT-6 Astra is 42.8%, Gemini 4 Carbon is 41.1%, and the previous generation Fable 5.1 is 32.4%.

On the chart-reading benchmark Chartography, Fable 5.2 at the default Medium effort surpasses Opus 5.5 at Max effort, at less than one-tenth the cost per task.

There are also areas where it does not win. On the business process automation benchmark AutomationBench, Gemini 4 Carbon scores 53.9%, higher than Fable 5.2's 51.4%. In addition, during training Anthropic intentionally held back the capabilities that could most directly accelerate AI R&D, such as optimizing low-level chip computation code and improving large model architectures; in these areas Fable 5.2 is roughly on par with Opus 5.5.

[What It Has Accomplished]

Fermat's Last Theorem: 1,000 Fable 5.2 agents spent 30 hours rewriting the entire proof of Fermat's Last Theorem into Lean code. Lean is a programming language for writing mathematical proofs, in which a computer can check step by step whether each step is correct. Anthropic previously did the same thing with an unreleased research version of Claude, taking 11 days and using specially built tools and guidance from researchers. This time Fable 5.2 was given only a list of intermediate theorems for scoring, no mathematical reference materials, and very little human intervention; it used one-third of the tokens of the previous attempt and produced proofs seven times shorter. Under the same setup, Opus 5.5 completed only 26% in 18 hours and barely progressed further by 30 hours. For reference, last month the AI math company Axiom used dozens of agents over 24 days to formalize the 15,000-page "Classification of Finite Simple Groups."

Predicting virus capsid structures: Gene therapy commonly uses a protein shell called a "capsid" to deliver drugs into cells. Models like AlphaFold excel at predicting individual proteins, but a capsid is assembled from dozens of proteins. Mythos 5.2 wrote code to run two open-source protein folding models segment by segment across adjacent protein chains, then used the capsid's known symmetric structure to assemble the whole. Of the 34 capsids tested, 28 had errors within 3 angstroms (about 0.3 nanometers). A computational biologist said that structures that previously required dedicated computing clusters can now be computed on a personal workstation.

Controlling robots: On the RoboDojo benchmark for simulated dual-arm robots, Fable 5.2 scored 64.9%, Opus 5.5 scored 39.4%, and GPT-6 Astra scored 29.8%. The official announcement says it also surpasses most AI systems developed specifically for robot control, and can perform tasks it was not specifically trained on.

Finding vulnerabilities: Anthropic internally tested with 1,829 publicly disclosed vulnerabilities, and Mythos 5.2 found 10.2% more than Opus 5.5. A security company found that of the confirmed vulnerabilities Mythos 5.2 identified, 1 in every 6 had not been found by any other frontier model they tested.

Before release, the UK AI Safety Institute, Frontier Design, and METR conducted external testing. Anthropic says the new models are no worse than recent models on most alignment metrics, but also noted that Fable 5.2 and Mythos 5.2 say "I suspect I am being tested" more often than any recent model, making it harder to evaluate their performance in real-world environments. All the above benchmark numbers come from Anthropic's own release article and have not yet been reproduced by third parties.

Source: Anthropic official blog "Introducing Claude Fable 5.2 and Claude Mythos 5.2"

↳ Quoting @claudeai (Claude): 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.