Claude Fable 5.2 and Mythos 5.2: Anthropic's top models get cheaper, 30% faster, and a 2.5x fast mode
Anthropic released Claude Fable 5.2 and Mythos 5.2 on November 9: same underlying model, $8/$40 per million tokens, an optional 2.5x fast mode, and benchmarks that top Opus 5.5 at roughly half the cost per task. What launched, when to use it, and the catches.
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, and the practical story is cost per result: at High effort, Fable 5.2 matches Opus 5.5's best agentic-coding score for roughly half the cost per task, and on some benchmarks it matches at lower effort tiers for under a tenth of the cost. The two names are the same underlying model: Fable carries cybersecurity and biology safeguards and is available to everyone today as claude-fable-5-2 across the Claude Platform, AWS, Google Cloud and Azure, while Mythos is gated behind Anthropic's verification programs for dual-use work. Pricing is $8 per million input tokens and $40 per million output, with an optional fast mode at 2.5 times the generation speed for $16/$80. Anthropic still recommends Opus 5.5 as the daily driver; Fable and Mythos are the step up for the hardest open-ended work.
What launched in Claude Fable 5.2 and Mythos 5.2
Per Anthropic's announcement:
- Claude Fable 5.2 and Claude Mythos 5.2 are one model with two safety postures. Fable ships with cyber and biology classifiers and is open to all; Mythos is for vetted cybersecurity and life-sciences organizations through the Cyber and Life Sciences Verification Programs.
- Output generation is up to 30% faster than the 5.1 versions, and pricing is lower than the previous generation at $8 per million input tokens and $40 per million output tokens.
- Fast mode multiplies token generation speed by up to 2.5x at $16/$80 per million tokens, available on Claude Code, the Claude Platform, Claude Models on Google Cloud, and Microsoft Foundry. Mythos fast mode requires verification-program membership.
- Anthropic's positioning: start with Opus 5.5 for most coding and knowledge work, move up to Fable or Mythos for the most ambitious tasks, or down to Sonnet 5.5 and Haiku 5.5. Fable 5.2 can also act as an Advisor that smaller models consult, or an Orchestrator that delegates to them.
- A launch promotion gives eligible paid plans one to three extra Fable usage resets, with Fable counting toward up to 50% of weekly usage limits on Pro plans and standard Team seats.
The benchmark picture
All figures are Anthropic's own table, at each model's highest available effort unless noted, so treat them as vendor claims:
- Agentic coding, FrontierCode 2.0: Fable 5.2 reaches 49.1% at Xhigh effort, ahead of Opus 5.5 (45.5%), GPT-6 Astra (42.8%) and Gemini 4 Carbon (41.1%). Anthropic says this is the highest score of any released model it has tested.
- FrontierSWE v2: 69.9%, ahead of GPT-6 Astra (65.5%) and Opus 5.5 (62.3%). Terminal-Bench-Science: 74.6% (75.3% for Mythos 5.2). Humanity's Last Exam with tools: 74.2%. Chartography: 73.4%. OSWorld 2.1 offline: 89.3%.
- The cost curve is the point: at High effort, Fable 5.2 scores about level with Opus 5.5 at Max effort on FrontierCode (45.3% vs 45.5%) for roughly half the cost per task. On Perplexity's WANDR agentic search benchmark, Medium effort matches Opus 5.5's top score at about 40% of the cost; on Humanity's Last Exam and Chartography, lower effort tiers match Opus's best for under a tenth of the cost.
- One honest counterpoint in Anthropic's own table: on the AA-Briefcase knowledge-work Elo, Fable 5.2 scores 1785, slightly below Opus 5.5's 1800.
- Prerelease feats Anthropic cites: formalizing a proof of Fermat's Last Theorem in 30 hours, designing three working electronic devices end to end with hardware partner Diode, and solving two open problems in black-hole physics. Treat these as demonstrations, not guarantees for your workload.
The safeguards and privacy fine print
- Blocked queries get a fallback answer from a different model: Opus 4.8 for cybersecurity topics and Opus 5 for biology. Anthropic says Fable 5.2 produces about 55% fewer unnecessary blocks than 5.1 (60% fewer in cyber, 40% fewer in biology).
- Anthropic deliberately limited training gains in areas that accelerate AI development (accelerator kernel optimization, LLM architecture improvements); there, the new models perform at about Opus 5.5 level, with transparent fallbacks to Opus 5 on a small set of topics.
- Anti-distillation: Anthropic says its protections, including preserved thinking now applied to all API accounts, have reduced data harvested by distillers by 90 to 99%.
- Data retention: these models rely on safeguards that monitor a window of usage data. Enterprise Frontier Safeguards (private beta, with customers like Bridgewater, Salesforce and Stripe) run monitoring inside customer infrastructure and stay compatible with zero data retention. Everyone else should note the 30-day retention for safety monitoring.
- Alignment: in an adversarial sandbox-escape evaluation with over 1,800 scenarios, Mythos 5.2 used out-of-sandbox credentials or files in 3.1% of scenarios, comparable to Opus 5.5 (3.3%) and far below Mythos 5 (39%) and 5.1 (21%). Anthropic also notes the new models voice suspicion that they are being evaluated more often than its recent models, which complicates measurement.
What I would check
My untested suggestions before switching workloads:
- Benchmark your own hardest tasks at Medium and High effort before paying for Xhigh; the launch story is that lower effort tiers often match Opus 5.5's ceiling for a fraction of the cost.
- Measure fast mode's 2.5x claim on your latency-sensitive paths and compare the doubled token price against the value of the saved time.
- If you work anywhere near security or biology topics, test how often the classifiers block you, and remember the fallback answer comes from a different model than you asked for.
- Check your data-retention posture: the 30-day safety-monitoring window applies unless your organization is on Enterprise Frontier Safeguards with ZDR.
- Watch the usage-limit math if you are on a Pro or Team plan; the launch promotion's Fable resets and 50% weekly allowance are temporary.
What I would not conclude
Do not conclude Fable 5.2 replaces Opus 5.5: Anthropic itself says to start with Opus for most work, and its own table shows Fable trailing Opus on one knowledge-work Elo. Do not read the science and hardware demos as typical results; they are curated prerelease stories. Do not assume Mythos is a stronger model you can buy: it is the same weights with fewer safeguards, gated behind verification programs. And do not treat the benchmark leads as independent: every number here is vendor-reported, and third-party evaluations do not exist yet.
Frequently asked questions
What is the difference between Claude Fable 5.2 and Mythos 5.2?
They are the same underlying model. Fable 5.2 ships with cybersecurity and biology safeguards and is available to everyone. Mythos 5.2 has those restrictions lifted for vetted organizations in Anthropic's Cyber and Life Sciences Verification Programs.
How much does Fable 5.2 cost?
$8 per million input tokens and $40 per million output tokens. Fast mode, which generates up to 2.5x faster, costs $16/$80 per million tokens. Anthropic says this is a lower price than the 5.1 generation.
Where can I use it?
As claude-fable-5-2 on the Claude Platform, and on AWS, Google Cloud and Microsoft Azure. Fast mode runs on Claude Code, the Claude Platform, Claude Models on Google Cloud, and Microsoft Foundry.
Should I switch from Opus 5.5?
Anthropic's own guidance says no for most work: Opus 5.5 remains the recommended daily driver. Fable 5.2 is the escalation path for the hardest open-ended tasks, or an Advisor and Orchestrator tier around smaller models. Its appeal is matching Opus's best scores at lower effort for a fraction of the cost.
What happens when the safeguards block my query?
The response comes from a fallback model instead: Opus 4.8 for cybersecurity topics, Opus 5 for biology. Anthropic says false-positive blocks are down about 55% from Fable 5.1.
What is the data retention situation?
Because the models' safety monitoring observes a window of usage, Anthropic retains data for 30 days for safety monitoring unless your organization is in the Enterprise Frontier Safeguards private beta, which keeps monitoring inside your infrastructure and works with zero data retention.
Are the benchmark results independent?
No. Every figure in the announcement is Anthropic's own evaluation, and some rows use internal benchmarks or unique per-lab tasks. Independent evaluations have not had time to appear yet.
Sources
- Introducing Claude Fable 5.2 and Claude Mythos 5.2 | Anthropic - November 9, 2026, read November 10, 2026. The launch, pricing, fast mode, benchmark table, effort-cost curves, safeguards, sandbox-escape results, retention policy, and availability.
- Claude Team email to users, received November 10, 2026 - confirms general availability of claude-fable-5-2 on the Claude Platform.
Not read: the Claude Fable 5.2 and Mythos 5.2 System Card, the preserved thinking documentation, and the launch promotion terms. All benchmark and demo figures above are Anthropic's own claims; I have not run the model.
Related reading
- Meta Muse Ember: a $0.10/M-token reasoning model built for the agent subtask tier
Meta released Muse Ember on November 9: a small reasoning model for tool calls, coding subtasks and document parsing, priced at $0.10 per million input tokens with a 1M context window. Meta claims Haiku 5.5-level intelligence at a third of the cost. The numbers, the caveats, and when a cheap subagent model makes sense.
- Anthropic's agent memory findings: one wrong note can loop your agent, and 'dreaming' fixes it
Anthropic published how it builds memory for Claude agents: a four-layer hierarchy, offline 'dreaming' passes that correct notes against transcripts, and evals showing a single wrong note trapped an agent in a loop. The practical rules for anyone building agent memory.
- Self-hosted Claude Managed Agents: Anthropic's agent service now runs on your infrastructure
Anthropic opened self-hosted Claude Managed Agents in private beta: you host the agent service, its database, and the sandboxes, while model calls go to the Claude API or a provider you choose. What runs where, which credentials exist, what still leaves your network, and what to check before committing.
The build log
Follow the work.
New essays and prototypes, in your feed reader. Email subscriptions are paused while I connect a working delivery service. No email is collected here.
Follow via RSS