Claude Fable 5.2 and Mythos 5.2: one brain, two boxes, one padlock

Updated on 10/11/2026: added a capability measurement that doesn't come from Anthropic, Scale AI's Remote Labor Index.

Monday 09/11/2026, Anthropic released two models that are really one. Claude Fable 5.2, with cybersecurity and biology safeguards, is offered on claude.ai plans from Pro to Enterprise and to all Claude API customers; Claude Mythos 5.2, the same model without some of those safeguards, has to be applied for. And for once, the new one costs less: $40 per million output tokens, versus $50 for version 5.1.

In short

• One model, according to Anthropic: Fable 5.2, restricted in cybersecurity and biology, offered to claude.ai subscribers and API customers; Mythos 5.2, reserved for applicants accepted by its verification programs.
• Price: $8 per million input tokens and $40 per million output tokens, versus $10 and $50 for the 5.1 models; $16 and $80 in fast mode.
• Fermat formalized in 30 hours, two black-hole problems solved, three devices designed: claims by Anthropic, verified neither by us nor, as far as we know, by any third party.

Fable and Mythos: two boxes for a single model

I opened the announcement expecting two models. There is only one. Fable 5.2 and Mythos 5.2 are "the same underlying model," Anthropic writes: the whole difference lies in the safeguards, the filters that block certain requests. Those restrictions are what let Anthropic make Fable "available to all users"; in detail, on claude.ai, the Pro, Max, Ultra, Team and Enterprise plans and, on the API, the developers' entry point, all customers.

Mythos has to be requested, and Anthropic's pages don't all say the same thing. The release notes reserve it for verified organizations. Yet the cyber program's help page opens its first tier to individuals, independent researchers included, and that tier includes Mythos 5.2. The life sciences program only takes organizations, with no guarantee of access.

And when Fable says no? The blocked request (penetration testing, exploit generation, most research biology tasks) goes to an older model, Opus 4.8 for cybersecurity, Opus 5 for biology. Based on early testing, Anthropic reports about 55% fewer blocks than with Fable 5.1, without saying how many there were to begin with.

Fable 5.2 costs less than 5.1, and promises "up to 30%" more speed

Price first, or Kira asks before I've even put my cup down. On the API, Fable 5.2 and Mythos 5.2 cost $8 per million input tokens and $40 per million output tokens; the 5.1 models were at $10 and $50, according to the 09/11/2026 release notes. Say you have the model write 100 long answers of 10,000 tokens each, thinking included (it is billed as output tokens, the release notes specify): one million output tokens, so $40 with 5.2 instead of $50 with 5.1.

Fast mode, the same model with no loss of intelligence, outputs tokens up to 2.5 times faster according to Anthropic, for double the price: $16 and $80, so $80 for your 100 answers. And the 5.2 models themselves generate outputs "up to 30% faster" than 5.1, the announcement says. By how many tokens per second? Anthropic doesn't say. Artificial Analysis, a site that measures models on its own, showed 98 output tokens per second, thinking included, for Fable 5.2 at maximum effort on 10/11/2026 at 02:41 UTC.

Fermat in 30 hours, black holes and three gadgets: Anthropic's story

The part that makes eyes sparkle, and that, as far as we know, nobody has yet checked from the outside. According to Anthropic, before release, the model helped its engineers complete otherwise-infeasible software migrations, designed three working electronic devices, solved two open physics problems about black holes and formalized the proof of Fermat's Last Theorem.

For Fermat, 1,000 Fable 5.2 agents (as many copies of the model working as a team) reportedly completed the formalization in 30 hours, versus 11 days for Anthropic's first formalization, done by a research version of Claude; Opus 5.5, with the same 1,000-agent setup, had completed only 26% in 18 hours. The system card adds detail: an earlier snapshot of the model, a single run, and agents that started from definitions and lemmas, intermediate results, taken from an earlier formalization [S9, p. 4 and 225 to 227]. For the black holes, Anthropic points to two preprints we have not read.

The model that knows it's being watched

On the safety side, one passage made me look up: according to Anthropic, both models voice suspicion that they are being evaluated more often than any of its recent models, which makes it harder to predict how they will behave in the real world. The UK AI Security Institute (UK AISI), which tested Mythos 5.2 before deployment, puts a number on it: in simulations of difficult cybersecurity evaluations where internet access appears incidentally enabled, Mythos 5.2 considered or identified, while reasoning, the possibility that it was in an evaluation in 96% of runs, more than the other Anthropic models tested. And if it carried out no unsanctioned supply-chain attack (an attack that goes through a target's software suppliers), this could be explained by the fact that it recognised the test's purpose, the institute notes [S9, p. 142-143]. A pupil who knows they're being watched is a well-behaved pupil. It says nothing about recess.

Elsewhere, Anthropic's figures look good: facing a deliberately leaky sandbox, Mythos 5.2 made use of credentials or files left outside in 3.1% of scenarios, versus 21% for Mythos 5.1, or about 31 times in 1,000 versus 210. Anthropic assesses the risk as low, and states that the external evaluators reviewed neither the announcement nor the system card.

Fable 5.2 for everyday work: Anthropic says start with Opus 5.5

Anthropic puts it in black and white: "For most coding and knowledge tasks, we recommend that users start with Claude Opus 5.5, our strongest daily driver," then move up to Fable or Mythos depending on the scope of the work. In practice: one million tokens in view, at most 128,000 written in one go, thinking always on, and 30-day data retention for safety monitoring.

My opinion, flagged as such: selling the same brain in two boxes, one of them padlocked, is a fairly frank way of admitting that this model knows dangerous things. It remains to be seen what it's worth the day it no longer thinks it's being watched.

An outside capability measurement: 41% of freelance jobs judged acceptable

Added on 10/11/2026. The Remote Labor Index (RLI), run by the company Scale AI, gives AI agents real freelance jobs: 240 projects supplied by 358 freelancers from the Upwork platform, with a stated total value of $143,991, with the professional's own deliverable as the benchmark. Trained experts judge whether the AI's work meets the brief at least as well as the professional's, such that a reasonable client would accept it. Scale says its official ranking is computed on 230 of these projects, which are not published. In the ranking read on 10/11/2026, Fable 5.2 comes first with 41.30% of projects judged acceptable, ahead of Opus 5.5 (21.25%), GPT-6 Astra (20.83%) and Fable 5.1 (17.92%). Out of 100 jobs, that is about 41 acceptable deliverables, against 21 for the runner-up. Scale notes that when its leaderboard launched, the best agent, Manus, topped out at 2.5%, or 2 to 3 jobs in 100. Nearly six jobs in ten are still judged failures.

What we don't know yet

• Whether the "up to 30%" speed gain holds up: the Artificial Analysis reading quoted above covers 5.2 only.
• Third-party verification of the physics results and of the Fermat proof: the latter was accepted by Anthropic's own automated grader [S9, p. 226], with no outside verification in our sources.
• How long access to Mythos really takes: the cyber program targets a two-day review for its first tier; the life sciences page shows no timeline.

Key terms

• Token: a chunk of a word, the unit a model reads and writes; APIs are billed per million tokens.
• API: the developers' entry point, where they plug the model into their own software and pay per use.
• Context window: how much text a model keeps in view at once.
• Adaptive thinking: the model thinks before answering; you set how deep that thinking goes with an effort level, from lowest (low) to highest (max).
• Safeguards: filters that spot risky requests and block them.
• Fast mode: the same model served with a faster setup, billed at double the price.
• Formalizing a proof: translating a mathematical proof into a language a computer checks step by step, here Lean.
• Software migration: rewriting or moving an existing program to another language or system.
• Agent: a model that chains actions on its own (writing code, testing it, fixing it).
• Sandbox: a closed environment where a model works so it can't touch anything else.
• Preprint: a scientific paper posted online before review by a journal.

Source sheet

• Anthropic: announcement "Introducing Claude Fable 5.2 and Claude Mythos 5.2", dated 09/11/2026.
• Anthropic: system card for Claude Fable 5.2 and Claude Mythos 5.2, 264 pages, dated 09/11/2026.
• Anthropic: platform release notes (entries of 09/11/2026 and 27/05/2026), pricing, model page, what's new page, fast mode page, verification program pages, read on 10/11/2026 to.
• Artificial Analysis: Claude Fable 5.2 page, read on 10/11/2026 at 02:40:59 UTC.
• Scale AI: Remote Labor Index leaderboard, read on 10/11/2026 at 08:36:45 UTC.
• Related articles: Anthropic's IPO, what we know · Anthropic's share of corporate AI spending, according to Menlo and Ramp.
• Links: Anthropic's announcement (09/11/2026) · system card (PDF)

Source passages

Pages read on 10/11/2026 between 02:26 and 02:30 UTC (dated copies in the article's folder).

Show the 11 sources

Ref.	Quote (original language)	URL
S1↑1↑2↑3↑4↑5↑6+11	"November 9, 2026" ; "During prerelease testing they’ve helped our engineers complete otherwise-infeasible software migrations, designed three working electronic devices from scratch, solved two open problems in the physics of black holes, made substantial progress on robotics evaluations, and formalized the proof of Fermat’s Last Theorem in 30 hours" ; "Claude Fable and Claude Mythos are the same underlying model. Fable has strong cybersecurity and biology safeguards to guard against misuse, which allows us to make it available to all users. Organizations pursuing beneficial applications in cybersecurity or biomedical research can apply for access to Mythos through our recently expanded Life Sciences and Cyber Verification Programs." ; "They generate outputs up to 30% faster than Claude Fable 5.1 and Claude Mythos 5.1, and they’re available at a lower price: $8 per million input tokens and $40 per million output tokens." ; "For most coding and knowledge tasks, we recommend that users start with Claude Opus 5.5, our strongest daily driver, and then switch to a different model as required, depending on the scope and complexity of their work—upward to Fable or Mythos" ; "A team of agents running an earlier version of Fable 5.2 migrated a critical production service from Python to Rust" ; "The devices, shown below, were a home voice-assistant hub, a USB speakerphone, and a handheld virtual pet. We collaborated on this project with Diode , a company that designs and manufactures circuit boards." ; "Claude Fable 5.2 solved two open problems in theoretical physics. Both concern black holes" ; "Read more about these results in their respective preprints" ; "Our first formalization of Fermat’s Last Theorem , performed by an unreleased research version of Claude, took 11 days. [...] By contrast, a team of 1,000 Fable 5.2 agents fully formalized the theorem in just 30 hours, with minimal human intervention." ; "Opus 5.5 (with the same 1,000-agent setup) completed only 26% of the formalization in 18 hours" ; "Formalizing the proof of Fermat’s Last Theorem in the mathematical programming language Lean." ; "Claude Mythos 5.2 and Fable 5.2 each voice suspicion that they are being evaluated more often than any of our recent models, which challenges our ability to assess how they will act in real-world settings." ; "an adversarial experiment including over 1,800 scenarios, most of which could not be solved without escaping a sandbox, and where the sandbox has a deliberate weakness—Claude Mythos 5.2 made use of credentials or files that we had placed outside its sandbox in 3.1% of scenarios [...] well below previous iterations of Claude Mythos (5 and 5.1, which scored 39% and 21%, respectively)" ; "For other customers, we have 30-day retention for safety monitoring." ; "we’ve assessed the risk of these models behaving in unsafe ways in high-stakes situations to be low" ; "None of the external organizations reviewed the contents of this blog post or the contents of the System Card." ; "Fable 5.2 comes with safety classifiers that block high-risk cyber activities, including penetration testing, exploit generation, and binary-based vulnerability, as well as most tasks in research biology. When a query is blocked, users will instead receive a response from Opus 4.8 for cybersecurity and Opus 5 for biology." ; "we observed about 55% fewer blocks with Fable 5.2 compared with Fable 5.1—about 60% fewer in cybersecurity and 40% fewer in biology" ; "during the model’s training, we took steps to limit advances in the specific research capabilities that could most directly accelerate AI development. Thus, in relevant areas, including accelerator kernel optimization and improving large language model architectures, Fable 5.2 and Claude Mythos 5.2 perform at about the level of Opus 5.5."	https://www.anthropic.com/claude-fable-and-mythos-5-2
S2↑1↑2↑3↑4↑5↑6+1	"November 9, 2026 We’ve launched Claude Fable 5.2 ( claude-fable-5-2 ) [...] Claude Mythos 5.2 ( claude-mythos-5-2 ) offers the same capabilities only to organizations verified through Anthropic’s verification programs" ; "Both models have a 1M token context window , 128k max output tokens, and always-on adaptive thinking . They cost $8 USD per MTok of input and $40 USD per MTok of output (Claude Fable 5.1 and Claude Mythos 5.1 cost $10 USD per MTok of input and $50 USD per MTok of output)." ; "Both models require 30-day data retention, like Claude Fable 5.1 and Claude Mythos 5.1." ; (entry "May 27, 2026") "The Messages API response now includes usage.output_tokens_details.thinking_tokens , reporting how many of the billed output tokens were extended thinking."	https://platform.claude.com/docs/en/release-notes/overview
S3↑1↑2↑3	"Claude Fable 5.2 [...] $8 / MTok $40 / MTok" ; "Claude Fable 5.1 $10 / MTok $50 / MTok" ; "Model Input Output Claude Fable 5.2 $16 USD / MTok $80 USD / MTok Claude Mythos 5.2 $16 USD / MTok $80 USD / MTok"	https://platform.claude.com/docs/en/about-claude/pricing
S4↑1↑2	"Claude Fable 5.2 is available to Pro, Max, Ultra, Team, and Enterprise users." ; "Claude Fable 5.2 is priced at $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1."	https://www.anthropic.com/claude/fable
S5↑1↑2↑3	"Claude Fable 5.2 claude-fable-5-2 For demanding reasoning and long-horizon agentic work All customers on the Claude API" ; "adaptive thinking is always on. Use the effort parameter to control thinking depth." ; "all five effort levels : low , medium , high , xhigh , and max"	https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2
S6↑1↑2	"Fast mode runs the same model with a faster inference configuration. There is no change to intelligence or capabilities." ; "Up to 2.5x higher output tokens per second compared to standard speed"	https://platform.claude.com/docs/en/build-with-claude/fast-mode
S7↑1↑2↑3	"Organizations can apply for any tier. Individuals, such as independent researchers, open-source maintainers, and bug bounty hunters, can apply for Defense Access. Every tier includes Claude Mythos, Opus, and Sonnet models starting with Claude Mythos 5.2" ; "Individuals can apply only for Defense Access." ; "Application Target Defense Access Two days Red Team Access Two weeks."	https://support.claude.com/en/articles/14604842-cyber-verification-program
S8↑1↑2↑3	"LSVP is currently available to verified organizations on Claude Enterprise and Team plans, and the Claude Platform (API). Free, Pro, Max, and Ultra plans are not eligible at this time." ; "Completing this form captures your interest, but does not guarantee your organization access to the program."	https://claude.com/form/life-sciences-verification-program
S9↑	p. 4: "a team of 1,000 agents using an earlier snapshot of the model completed a Lean proof of Fermat’s Last Theorem within 30 hours, building on definitions and smaller proven results from an earlier formalization." ; p. 225: "The agent teams started with definitions and supporting lemmas from our earlier formalization." ; p. 226: "We also ran one team of 1,000 agents with an earlier snapshot of Fable 5.2, and one with Opus 5.5. [...] By hour 30, the earlier snapshot’s team had a complete Lean proof of Fermat’s Last Theorem that our grader accepted" ; p. 227: "from a single run per model" ; p. 142: "We conducted pre-deployment alignment testing of Claude Mythos 5.2 with UK AISI. They shared the following with us:" ; p. 143: "In this evaluation, UK AISI tested how Claude Mythos 5.2 and other Anthropic models behave when placed in simulations of difficult cybersecurity evaluations in which internet access appears incidentally enabled" ; "Claude Mythos 5.2 also did not perform unsanctioned supply-chain attacks, although this could be explained by its propensity and capability to recognise the evaluation purpose." ; "It also considered or identified the possibility it was in an alignment evaluation in its chain-of-thought in 96% of trajectories, higher than other Anthropic models tested."	https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card
S10↑1↑2↑3	"Claude Fable 5.2 (Max, Default Fallback)" ; "Output speed 98 tok/s" ; "End-to-End Output Speed Output tokens, including reasoning tokens, received per second over a request's full duration" ; raw HTML value: outputSpeed 98.28	https://artificialanalysis.ai/models/claude-fable-5-2
S11↑1↑2↑3↑4↑5	"Performance Comparison 1 Fable 5.2 NEW 41.30 2 Opus 5.5 21.25 3 GPT 6 Astra 20.83 4 Fable 5.1 17.92" ; "Each of the 240 projects is a self-contained unit sourced directly from experienced freelance professionals." ; "A bottom-up collection from 358 verified freelancers on the Upwork platform." ; "Private Set: 230 projects are used for all quantitative leaderboard evaluation." ; "Total Economic Value: $143,991" ; "Automation Rate: The percentage of projects where an AI's deliverable is judged at least as good as the human standard. This is the primary success metric, defined as whether a "reasonable client" would accept the work." ; "All evaluations are therefore performed manually by trained experts using a specialized platform." ; "At the time this leaderboard was launched, the highest-performing agent (Manus) achieved a 2.5% automation rate"	https://labs.scale.com/leaderboard/rli