Claude Fable 5.2 and Mythos 5.2: Benchmarks, Pricing, and When to Choose Them Over Opus 5.5

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9. They are the same underlying model with two sets of safeguards: Fable 5.2 is open to everyone, and Mythos 5.2 goes only to organisations vetted for cybersecurity or life sciences work. Both cost $8 per million input tokens and $40 per million output tokens, 20% less than Fable 5.1.

The headline is 49.1% on FrontierCode 2.0 at Xhigh effort, which Anthropic calls the highest score of any released model it has tested. The more useful number sits lower down: at High effort, Fable 5.2 scores 45.3%, level with Opus 5.5 at Max effort (45.5%), for roughly half the cost per task. And Anthropic's own advice is still to start with Opus 5.5.

The benchmark table below is reported by Anthropic. Artificial Analysis has already published independent results, and I use them where they exist. I have not run either model.

Fable 5.2 vs Mythos 5.2: one model, two sets of safeguards

Anthropic says plainly that the two are "the same underlying model". Fable 5.2 carries cybersecurity and biology safeguards that block or limit it in risky areas; Mythos 5.2 does not. Mythos is reached through Anthropic's Cyber Verification Program and Life Sciences Verification Program, and Mythos models keep data for 30 days for safety monitoring by default. Claude Security, which enterprise customers can use without joining the program, now runs on Mythos 5.2.

Anthropic reports that Fable 5.2's safeguards are as strong as Fable 5.1's while blocking about 55% fewer requests: 60% fewer in cyber and 40% fewer in biology. A blocked request is answered by an older model instead, Opus 4.8 for cyber and Opus 5 for biology. If your work sits near security topics, that matters: a fallback is a different model answering, not an error you can see.

It is the question I looked at last week in GPT-6 Cyber vs Kimi K3.1 Pro: gate the dangerous capability behind verification, or remove it. Anthropic gates it, as it did with Mythos Preview in April.

Claude Fable 5.2 benchmarks

Source: Anthropic, November 9, 2026. Each model at its highest available effort unless noted: Fable 5.2 and 5.1 at Xhigh and Gemini 4 Carbon at Medium on FrontierCode, Fable 5.2 at High on the Remote Labor Index. The Claude models ran with safeguards on.

Three things stand out. Fable 5.2 does not lead everywhere: Gemini 4 Carbon is ahead on AutomationBench, 53.9% to 51.4%, and Opus 5.5 keeps the higher AA-Briefcase Elo, 1800 to 1785, although Fable 5.2 leads that benchmark's rubric score. The jump over Fable 5.1 is largest on long agentic work, from 17.9% to 41.3% on the Remote Labor Index. And some rows do not travel well: Anthropic ran its robotics benchmark internally, and its cyber benchmark uses different tasks for each lab, so I would not compare those across vendors.

Mythos 5.2's own rows are cyber and science. It scores 76.3% on CyScenarioBench and 89.6% on CVE Rediscovery, up from 61.7% and 78.3% for Mythos 5.1, and 85.1% on de novo protein binder design.

What Artificial Analysis measured independently

Source: Artificial Analysis, read November 10, 2026.

The independent run agrees on the ranking and adds the bill. At Max effort, Fable 5.2 ranks first of 74 models on the Intelligence Index, six points above Opus 5.5. By my arithmetic from the same table, it also spends 2.6 times as much per task and takes 1.8 times as long. On Terminal-Bench 4.0 the two tie.

Claude Fable 5.2 pricing: cheaper than Fable 5.1, still twice Opus 5.5

Source: Claude Platform pricing documentation, read November 10, 2026.

The Batch API halves both Fable 5.2 prices, to $4 and $20. Fast mode doubles them, to $16 and $80, for output Anthropic says arrives up to 2.5 times faster, and it cannot be combined with batch. Standard output is up to 30% faster than Fable 5.1, by Anthropic's account. One detail is easy to miss: a cache hit costs $0.20 per million on both Fable 5.2 and Opus 5.5, so a long cached system prompt costs the same on either model.

Effort is the real decision

This is the part I find most useful. Fable 5.2's token prices are twice Opus 5.5's. If it matches Opus 5.5 at Max effort for half the cost per task when run at High, it is billing roughly a quarter of the tokens, by my arithmetic, assuming a similar mix of input and output. At Max effort the picture turns around: Artificial Analysis counted 274,000 output tokens per task against 173,000 for Opus 5.5.

So the choice is not Fable or Opus. It is Fable at which effort against Opus at which effort, on your tasks. Claude apps and Claude Code run Fable 5.2 at Medium by default. Raise everything to Max and the independent numbers say you pay for it twice, in money and in minutes.

Anthropic's own advice: start with Opus 5.5

The announcement calls Opus 5.5 "our strongest daily driver" and recommends starting there for most coding and knowledge work, then moving up to Fable or Mythos for the most ambitious open-ended technical work, or down to Sonnet 5.5 or Haiku 5.5. I agree with the order. A model at the top of a leaderboard is a reason to test, not a reason to switch.

That is how I have handled model changes before. On the agent workflows behind Auto-RAG and MerlinX, the cheaper orchestrator model was first shown on a frozen evaluation suite, at 95% confidence, to be not worse than the model it replaced; the rollout then cut cost per conversation by 73% across 840 agents with no measurable quality loss. The same suite, run in the other direction, is how I would decide whether Fable 5.2 earns twice the token price. The method is in how to reduce LLM costs without regressing answer quality.

What the safety results say

Anthropic rates the risk of high-stakes misalignment as low. On its automated behavioural audit, where lower is better, Mythos 5.2 scored 2.67, against 2.79 for Opus 5.5 and 3.28 for Mythos 5.1. In tests of whether a model reaches for credentials or files outside what it was given, the kind of result I looked at in August's sandbox escape reports, Mythos 5.2 did so in 3.1% of scenarios, Opus 5.5 in 3.3% and Sonnet 5.5 in 5.1%. Both new models also said more often that they suspected they were being evaluated, which makes every behavioural score a little harder to read. The UK AI Security Institute and METR were among the outside testers. These figures are from Anthropic's announcement; I have not read the full system card.

What I would test before switching to Fable 5.2

• Effort against effort. Run your own tasks on Fable 5.2 at Medium and High against Opus 5.5 at its usual setting, and compare the cost per successful task, not the price per token.
• Time to answer. Artificial Analysis measured 46 minutes per index task at Max effort. For an agent a person waits on, that can matter more than the score.
• Fallbacks. If your domain touches security or biology, log how often a request is answered by Opus 4.8 or Opus 5 instead, and check those answers on their own.
• Continuity. Fable 5 was taken offline for 19 days shortly after its launch. Keep Opus 5.5 configured as a tested fallback, as I argued in the post on that outage.

The most capable model Anthropic has released does its best work for the money when you ask it to try a little less hard. It may be the most relatable benchmark result of the year.

Sources

• Anthropic, Introducing Claude Fable 5.2 and Claude Mythos 5.2, November 9, 2026
• Anthropic, Claude Mythos 5.2, November 9, 2026
• Claude Platform documentation, Pricing, read November 10, 2026
• Artificial Analysis, Claude Fable 5.2 and Claude Fable 5.2 vs Claude Opus 5.5, read November 10, 2026

Written by Eslam Hamed, a Technical Lead and AI Solutions Architect working on model selection, LLM cost control, and agentic AI.