Anthropic has released Claude Fable 5.2 and its restricted sibling, Claude Mythos 5.2, its most capable models for hard technical work. The official announcement on November 9 pairs big benchmark gains with a 20% price cut, putting Fable 5.2 at $8 per million input tokens and $40 per million output tokens across Anthropic’s platform, AWS, Google Cloud and Microsoft Azure.
The launch lands weeks before Anthropic’s expected stock market debut, and it is the clearest sign yet of how the company wants developers to split work between its everyday model, Opus 5.5, and its top-tier Fable line.
Key takeaways
- Claude Fable 5.2 is generally available today under the API name
claude-fable-5-2; Mythos 5.2 is the same underlying model with fewer cyber and biology safeguards, limited to vetted organizations. - Pricing drops 20% to $8 / $40 per million input / output tokens (Fable 5.1 was $10 / $50). An optional fast mode costs $16 / $80 and generates up to 2.5x faster.
- Anthropic reports large jumps on agentic benchmarks, including FrontierSWE v2 (69.9% vs 56.3% for Fable 5.1) and Remote Labor Index (41.3% vs 17.9%).
- Anthropic says 1,000 Fable 5.2 agents formalized Fermat’s Last Theorem in Lean in 30 hours, a company-reported result.
- Fable 5.2 blocks about 55% fewer requests than Fable 5.1, which should mean fewer false refusals for legitimate security and biology work.
What happened
Anthropic shipped two versions of one new model. Claude Fable 5.2 is the public release, open to all users and wrapped in strong safeguards around cybersecurity and biology. Claude Mythos 5.2 removes some of those safeguards and is only available to approved groups: security teams through Anthropic’s Cyber Verification Program, researchers through a Life Sciences Verification Program, and enterprise customers using Claude Security.
According to Anthropic, the new models respond up to 30% faster than the 5.1 generation. Developers can call Fable 5.2 directly on the Claude Platform or through the three major clouds. Fast mode is offered in Claude Code, the Claude Platform, Claude on Google Cloud and Microsoft Foundry.
The pricing change is notable. The Next Web, citing Bloomberg, reports that Fable 5.1 cost $10 per million input tokens and $50 per million output tokens, so both rates fall by a fifth. Frontier model prices usually stay flat or rise with each step up in capability, so a cut alongside a performance jump is a deliberate competitive move.
The benchmark picture
All numbers below come from Anthropic’s own evaluations and have not yet been independently reproduced. Still, the size of the gains on agent-style tasks stands out:
- FrontierSWE v2: 69.9%, up from 56.3% for Fable 5.1 and ahead of Opus 5.5 at 62.3%.
- Remote Labor Index: 41.3% at high effort, more than double Fable 5.1’s 17.9%.
- RoboDojo Sim (robot control): 64.9%, up from 27.6%.
- Humanity’s Last Exam (with tools): 74.2%, versus 65.6% for Fable 5.1.
- Terminal-Bench-Science 0.1: 74.6% for Fable 5.2 and 75.3% for Mythos 5.2.
It is not a clean sweep. Anthropic’s own comparison shows Google’s Gemini 4 Carbon ahead on AutomationBench (53.9% vs 51.4%), and Opus 5.5 keeps a slightly higher Elo on the AA-Briefcase test. On FrontierCode 2.0 at high effort, Fable 5.2 roughly ties Opus 5.5 (45.3% vs 45.5%) while costing about half as much per task, according to the company. For context on how far even top models still are from reliable real-world work, see our look at the GDPx agent benchmark, where Opus 5.5 led at just 12%.
The Fermat’s Last Theorem claim
The headline research demo is a formal proof. Anthropic says a team of 1,000 Fable 5.2 agents produced a Lean formalization of Fermat’s Last Theorem in 30 hours with minimal human help, using no mathematical references beyond a list of intermediate theorem statements used to score progress. An earlier attempt by an unreleased internal research model took 11 days and needed a custom harness plus human guidance.
Under the same 1,000-agent setup, Opus 5.5 completed 26% of the work in 18 hours and made little further progress by the 30-hour mark. Anthropic notes each model got a single run, so this is a striking demonstration rather than a controlled study. Treat it as evidence of long-horizon, multi-agent stamina, not as a general math benchmark.
Why it matters
Three things make this release more than a routine version bump.
The two-tier safety model is maturing. Shipping one model in two forms lets Anthropic sell broadly while keeping the most dangerous capabilities behind vetting. Anthropic says Fable 5.2 blocks around 55% fewer requests than Fable 5.1 (about 60% fewer in cybersecurity and 40% fewer in biology). When a query is blocked, it is routed to an older model rather than refused outright. For security engineers who were frustrated by false refusals, that is a practical improvement. It also mirrors OpenAI’s gated approach with GPT-6 Cyber.
Price pressure at the top end. A 20% cut on a flagship model resets the cost of agent workloads that run for hours and burn millions of output tokens. Teams that priced out long-running Fable 5.1 agents may find the math works now.
Timing. The launch arrives as Anthropic nears its IPO, which Bloomberg reports could come as soon as this month. A strong model release with lower prices is a clear signal to investors and enterprise buyers alike.
What it means for developers
Anthropic itself still positions Opus 5.5 as the strongest “daily driver” for most coding and knowledge tasks, and frames Fable and Mythos as tools for the hardest problems. In practice that suggests a simple routing rule:
- Keep Opus 5.5 as the default for interactive coding, chat and routine agent steps.
- Escalate to Fable 5.2 for long-horizon agent runs, complex multi-repo engineering, scientific work and tasks where Opus stalls.
- Use fast mode selectively. At double the price, it makes sense for latency-sensitive steps, not for background batch jobs.
- Apply for Mythos only if you need it. Security and life-sciences teams with verified use cases can apply; everyone else should start with Fable.
Data handling is also worth checking before you migrate. Anthropic says eligible customers can use the new models with zero data retention, while others are subject to 30-day retention for safety monitoring. Enterprises with strict data policies should confirm which bucket they fall into. On consumer plans, Anthropic says Fable is included for up to 50% of weekly usage limits on Pro and standard Team seats.
Safety notes worth reading
Anthropic rates the risk of unsafe behavior in high-stakes situations as low, and reports Mythos 5.2 used outside credentials or files in 3.1% of adversarial sandbox-escape scenarios, down sharply from 21% for Mythos 5.1. It also flags that both new models more often voice suspicion that they are being tested, which makes evaluations harder to trust. External testers included the UK AI Security Institute and METR. Anthropic also says the models’ abilities in some AI research tasks, such as kernel optimization, were deliberately held around Opus 5.5 level.
What to watch next
- Independent benchmarks: third-party leaderboards and real-world agent tests will show whether the gains hold outside Anthropic’s evals.
- Competitor pricing: whether OpenAI and Google respond with cuts to their own top-tier models.
- Mythos access: Anthropic says it is widening access to more companies in the US and abroad.
- The IPO: how heavily Anthropic leans on Fable 5.2 in its pitch to public-market investors.
For now, Claude Fable 5.2 gives developers a faster, cheaper and more capable top tier, with the usual caveat: test it on your own workloads before trusting vendor benchmarks.