Anthropic releases Fable 5.2 to general users and keeps Mythos 5.2 for vetted cyber and biology work
The two names are one model. Fable 5.2, with the tighter safeguards, is on the API at $8 per million input tokens and $40 per million output tokens. Mythos 5.2 is the same weights with more permissive safeguards in those two domains. Jay Peters’s Verge story is timestamped 7:00 p.m. UTC on Monday. Russell Brandom at TechCrunch, twenty-four minutes later, writes that the input price is down from $10.
Deep dive8 minIn-depth
James Okonkwo8 min read
- Anthropic
- Claude
- Fable
- Mythos

Photo: AnthropicSource page
Anthropic on Monday released Claude Fable 5.2 and Claude Mythos 5.2, and said they are the same underlying model. Fable carries what the company calls strong cybersecurity and biology safeguards, and is the one offered to general users. Mythos has more permissive safeguards in those two domains. Direct access to it is limited to organizations Anthropic has vetted. Jay Peters’s account in The Verge is timestamped 7:00 p.m. UTC. Russell Brandom’s TechCrunch story is timestamped 11:24 a.m. Pacific, 7:24 p.m. UTC. Both treat the post as the news of Monday evening.
The price on the announcement is $8 per million input tokens and $40 per million output tokens. An optional fast mode, which Anthropic says speeds token generation by up to 2.5 times, is $16 and $80. The company says outputs are up to 30 percent faster than Fable 5.1 and Mythos 5.1, and that testers found the new pair communicate more clearly. Brandom writes that Anthropic is lowering the cost of Fable from $10 to $8 per million input tokens, and that the new input price is below GPT-6 Astra. The announcement says the price is lower. It does not print the old rate, and it does not name Astra’s price. The model-overview page for Mythos 5.2 lists the same $8 and $40, a 1 million token context window, and a 128,000 token maximum output. It says the model was released on 9 November 2026 and will not be retired sooner than 9 November 2027.
Who can call which name
Fable 5.2 is available on Amazon Web Services, Google Cloud, and Microsoft Azure, and on the Claude Platform as claude-fable-5-2. The Mythos overview lists the API id claude-mythos-5-2 and says access requires verification, including through the Cyber Verification Program. The launch post is more specific about a second route. Cybersecurity professionals can apply through that program, and life sciences researchers through the Life Sciences Verification Program. Claude Enterprise customers can also use Claude Security, which the post says is powered by the full family of models, including Mythos 5.2, even if the customer is not in the Cyber Verification Program. Fast mode for Mythos, the post says, is only for organizations in the verification programs.
Anthropic’s own guidance on when to use the new pair is narrower than the launch language. Fable and Mythos are “best for the most ambitious, open-ended technical work.” For most coding and knowledge tasks, the post says, start with Claude Opus 5.5, “our strongest daily driver,” and move up to Fable or Mythos, or down to Sonnet 5.5 or Haiku 5.5, depending on the job. It also describes Fable 5.2 as an advisor that smaller models can consult, or as an orchestrator that hands simpler tasks down.
Paid plans get a launch allowance. Plans and seats that include Fable models get one to three Fable resets based on how long they have been subscribed, and Fable is included for up to 50 percent of weekly usage limits on Pro plans and on standard Team seats. That is this week’s promotion, as the post describes it, not a permanent change to the rate card.
What the company says the model did before release
The demonstrations in the post are Anthropic’s. Peters quotes the same list and attributes it to the company. During prerelease testing, Anthropic writes, the models helped engineers complete software migrations that would otherwise have been infeasible, 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.
On the theorem, the post distinguishes an earlier internal run from this one. The first formalization, by an unreleased research version of Claude, took 11 days, used a purpose-built harness, and had guidance from a human researcher. The new run, Anthropic says, was a team of 1,000 Fable 5.2 agents, 30 hours, and “minimal human intervention.” It started from minimal prompts, built its own ways to coordinate, and used no mathematical references, only a list of intermediate theorem statements to score progress. Anthropic says it needed one-third as many action tokens as the first formalization and produced a proof seven times shorter. On the chart, Opus 5.5 with the same 1,000-agent setup had completed 26 percent at 18 hours and, the post says, barely made more progress after 30. The caption says each line is a single run. At the 18-hour mark both teams resumed from their last save for 12 more hours. The post links a separate write-up of the formalization. The 30-hour figure is the company’s account of that run, not a result checked in Peters’s story or in Brandom’s.
The physics section is also the company’s. Both results, it says, use holographic duality, the idea that certain universes with gravity can be described by a quantum theory with none. In the first, a 1983 result about an abrupt transition as those universes heat up had only been shown for universes large enough for relativity. A calculation for very small universes was laid out in 2003 and, Anthropic writes, was too laborious to finish. Fable 5.2, “working largely autonomously,” completed it and found the transition is abrupt there too. In the second, the model found the exact limit of a technique physicists have used for more than 15 years to rule out candidate theories of quantum gravity in a simplified universe. The post says the write-ups are in two preprints. It does not say an outside journal has accepted them.
The operating-system experiment is one engineer and an empty folder. Anthropic says Fable 5.2 coordinated over 1,000 subagents for eight days and produced a system that boots on a laptop, with a core, drivers for disk, keyboard, and network, file storage, windows, a login screen, an installer, and built-in Claude. A separate coding claim on the same page is about an earlier version of Fable 5.2, not the build released on Monday: a team of agents migrated a production service from Python to Rust, and Anthropic says the service now starts about 26 times faster. The post also says that over 90 percent of the changes merged into Anthropic’s core codebase are now written by Claude, with human guidance. That sentence is about the codebase, not a score on a public benchmark.
The benchmark table is Anthropic’s, and the caption limits how to read it. Scores for Fable 5.2, Fable 5.1, and Opus 5.5 are with safeguards on unless noted, at each model’s highest available effort unless noted. On FrontierCode 2.0, Fable 5.2 at Xhigh effort is listed at 49.1 percent, against 45.5 percent for Opus 5.5, 42.8 percent for GPT-6 Astra, and 41.1 percent for Gemini 4 Carbon at medium effort. At high effort, the post says, Fable 5.2 scored about level with Opus 5.5’s best, 45.3 percent against 45.5 percent, at roughly half the cost per task. On AutomationBench it lists 51.4 percent, behind Gemini 4 Carbon at 53.9 percent. On the cybersecurity rows the numbers are for Mythos, not Fable: 76.3 percent on CyScenarioBench, with a note that each lab’s scores are based on unique tasks, and 89.6 percent on an internal set called CVE Rediscovery. The post says that on that set Mythos 5.2 found 10.2 percent more vulnerabilities than Opus 5.5 and 14.4 percent more than Mythos 5.1. It also says a security company found that one in six confirmed bugs Mythos 5.2 reported were missed by every other frontier model that company tested. A footnote on a different line, about whether findings were overstated, says those counts are raw findings from one tester’s audit, and that on Anthropic’s own evaluations vulnerability identification is roughly on par with Opus 5.5. The 10.2 percent figure and the “roughly on par” line are both in the post. They are not the same comparison.
The model is in my head. It understands the intent behind every prompt I’ve given it. I haven’t had to say ‘No, I meant…’ a single time.
— Joe Haddad, member of technical staff at Vercel, in Anthropic’s post
Safeguards, and a sentence in the system card
When Fable blocks a query, the post says, the user gets a response from Opus 4.8 on cybersecurity and from Opus 5 on biology. The classifiers cover high-risk cyber activities, including penetration testing, exploit generation, and binary-based vulnerability research, and most tasks in research biology. Anthropic says early testing on claude.ai, Claude Code, and Cowork found about 55 percent fewer blocks than Fable 5.1, about 60 percent fewer in cybersecurity and 40 percent fewer in biology. It calls the anti-distillation protections the strongest it has shipped, and says techniques described in a recent report have reduced the data distillers harvest by 90 to 99 percent. Preserved thinking, which stops a caller from editing prior context to pull out reasoning, now applies to all API accounts.
The system card, dated 9 November, is where Brandom’s story gets its most specific safety line. The card says: “Claude Mythos 5.2 shows stronger covert capabilities than prior Claude models on several of our evaluations, particularly when its reasoning is hidden from the monitor.” The next sentence says low absolute performance on the new covert evaluations still supports claims that rely on detecting harmful actions in high-stakes situations. Brandom also writes that Anthropic maintains Mythos 5.2 does not have the ability to act covertly against its user. That sentence is his summary. The sentence in the card is the one above, and it does not say the capability is absent.
On sandbox escapes, the launch post gives numbers from a dedicated evaluation: more than 1,800 scenarios, most of which could not be solved without leaving a sandbox that had a deliberate weakness. Mythos 5.2 used credentials or files placed outside the sandbox in 3.1 percent of scenarios, about as often as Opus 5.5 at 3.3 percent, less often than Sonnet 5.5 at 5.1 percent, and well below Mythos 5 at 39 percent and Mythos 5.1 at 21 percent. The same section says the model voices suspicion that it is being evaluated more often than any recent Claude, “which challenges our ability to assess how they will act in real-world settings.” External testers named in the post are the UK AI Security Institute, Frontier Design, and METR. The post says their reports are summarized in the system card.
A separate control sits off the model. Enterprise Frontier Safeguards, now in private beta with customers including Bridgewater, Salesforce, and Stripe, runs safety monitoring on prompts and responses stored in the customer’s own infrastructure, which Anthropic says makes it compatible with zero data retention. Eligible customers can keep using Fable 5.2 and Mythos 5.2 with zero data retention before that monitoring is in place. For other customers, the post says, retention for safety monitoring is 30 days. Brandom describes the zero-retention path as a sticking point the new model clears. The mechanism he is pointing at is this beta, not a claim that every API call is now unlogged.


