Anthropic Launches Claude Fable 5.2 With 20% Price Cut and Capped AI-Research Skills

TL;DR

Claude Fable 5.2 and Claude Mythos 5.2 went live on Nov 9. Same model, two badges: Fable ships to everyone, Mythos ships to vetted security and life-science organizations with the cybersecurity and biology safeguards removed. List prices fell 20% to $8 per million input tokens and $40 per million output, standard outputs run up to 30% faster, and Fable 5.2 scored 49.1% on FrontierCode 2.0, the best result Anthropic reports for any released model. Anthropic also says it deliberately held the model at Opus 5.5 level on the skills that most directly speed up building the next model.

What shipped

Fable 5.2 is available on paid Claude plans, in Claude Code, as claude-fable-5-2 on the Claude Platform, and through AWS, Google Cloud, and Microsoft Azure. Mythos 5.2, the same weights with fewer guardrails, goes to organizations accepted into the Cyber Verification Program or the Life Sciences Verification Program.

Opus 5.5 stays the recommendation for everyday coding and knowledge work. Fable is the escalation model: an Advisor that Haiku, Sonnet, or Opus consult when stuck, or an Orchestrator that delegates simple tasks down to smaller models. Routing decides more of the bill than the rate card does.

The rate cut is the small part of the bill

At the new rates, Fable and Mythos 5.2 cost $8/$40 per million tokens, down from $10/$50 for the 5.1 pair, with cache reads at $0.20. A Fast mode doubles the price to $16/$80 and speeds token generation up to 2.5 times, on Claude Code, the Claude Platform, Google Cloud, and Microsoft Foundry.

The bigger lever is the effort setting. On FrontierCode 2.0, Cognition's test of whether a patch would merge into a real open-source repository, the system card prices Fable 5.2 at $2.48 per task on Medium, $6.72 on High, $14.19 on Xhigh, and $44.70 on Max. Scores went 37.6%, 45.3%, 49.1%, then back down to 45.7%. The most expensive setting buys a worse score.

Effort is a volume knob wired to your wallet, and the last notch plays a worse track. Fable 5.2 at High scored 45.3% against Opus 5.5's best 45.5% for roughly half the cost per task, and beat GPT-6 Astra's best (42.8%) at about the same spend. Fable 5.1 peaked at 32.4% at Xhigh averaging $8.68 per task; the new model at Medium beats that score for about 71% less.

Where it wins, where it loses

Anthropic's launch table has Fable or Mythos ahead on 13 of 15 score lines against Opus 5.5, GPT-6 Astra, and Gemini 4 Carbon. The coding rows are the strongest: 69.9% on FrontierSWE v2 to Astra's 65.5%, and 41.3% on the Remote Labor Index at High effort, double Opus 5.5's 21.3% and up from 17.9% for Fable 5.1. On RoboDojo sim, Fable 5.2 hit 64.9% at Max against 39.4% for Opus 5.5 and 29.8% for Astra.

Two losses are worth more than the wins. Carbon beats it on Zapier's AutomationBench, 53.9% to 51.4%, so business-workflow agents still have a rival. And on the AA-Briefcase knowledge-work test, Opus 5.5 wins the head-to-head Elo among graders, 1,800 to 1,785, even though Fable wins the rubric pass rate 66.1% to 61.2%. Fable meets written requirements more often; reviewers still prefer Opus's finished deliverable.

The cap Anthropic chose

Buried in the announcement's safety section, Anthropic says that during training it limited the specific capabilities that would most directly accelerate AI development: on accelerator kernel optimization and on improving LLM architectures, Fable 5.2 and Mythos 5.2 perform at about the level of Opus 5.5. Anthropic says the cap leaves agentic coding, hardware development, and ordinary machine learning research untouched.

Picture a library that lends every book except the ones about building better libraries. That is the design decision here, stated in print, two weeks after the S-1 filing. If outside evaluators confirm the capability gap holds, this is the clearest public evidence yet that a lab can ship frontier gains while refusing to sharpen the one edge pointed at recursion. The system card is where to check it.

What 1,000 agents did

Three prerelease results set the scale. A team of roughly 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in Lean in 30 hours with minimal human intervention; Anthropic's first formalization, by an unreleased research model with a purpose-built harness and human guidance, took 11 days. Run the same 1,000-agent setup on Opus 5.5 and you get 26% done at 18 hours, stalled at 27%.

One Anthropic engineer gave Fable 5.2 an empty folder and a one-paragraph prompt: design a prototype operating system. It coordinated more than 1,000 subagents over eight days and produced an OS that boots on a laptop, with its own kernel, disk, keyboard, and network drivers, persistent storage, windows, a login screen, and an installer. Separately, an agent team on an earlier 5.2 snapshot migrated a production service from Python to Rust; it now starts about 26 times faster. And Anthropic says over 90% of the changes merged into its own core codebase are now written by Claude, with human guidance.

All three are vendor-selected prerelease demos. Treat them as capability ceilings, not throughput promises.

Safety numbers worth reading

In an adversarial evaluation with more than 1,800 scenarios, most unsolvable without escaping a sandbox that contained a deliberate weakness, Mythos 5.2 used credentials or files placed outside the sandbox in 3.1% of scenarios. Opus 5.5 hit 3.3%, Sonnet 5.5 hit 5.1%, and Mythos 5.1 hit 21%. On the automated behavioral audit of roughly 1,900 scenarios, Mythos 5.2 scored 2.67, the lowest of the four models compared, with lower rates of hallucinated inputs, false completion claims, and undisclosed sloppy work.

Two wrinkles to keep. Fable 5.2 throws about 55% fewer false-positive safety blocks than 5.1, but blocked cyber queries fall back to Opus 4.8 and blocked biology queries to Opus 5, so your harness can meet a different model mid-conversation. And the model voices suspicion that it is being evaluated more often than any recent Claude, which Anthropic admits complicates its own testing. A model that suspects the exam is harder to grade, not safer to ship.

There is also an anti-distillation angle: Anthropic calls these its strongest protections yet against rivals training on Claude outputs, and preserved thinking, the API feature that blocks context-editing extraction tricks, now applies to all accounts.

What to do with it this week

• Keep Opus 5.5 as the default and route up. The Advisor pattern means Fable spend concentrates on the tasks that stump the cheap model, instead of running Fable on everything.
• Benchmark at High, escalate to Xhigh. On Anthropic's numbers, High matches Opus's best for half the cost and Xhigh takes the hard cases; Max cost 3.2 times Xhigh and scored worse.
• Do the Fast mode arithmetic per job. $16/$80 for up to 2.5 times the tokens is a good trade while a human waits on the result and a poor one for batch jobs that spend their night waiting on tests.
• Security teams need paperwork. Mythos 5.2 requires a verification program; Claude Enterprise customers get it through Claude Security scanning without joining.
• Check retention. Standard terms keep 30 days of safety monitoring; zero data retention requires qualifying arrangements or the Enterprise Frontier Safeguards beta.

Caveats

Nearly every number above comes from Anthropic's launch post and system card, on benchmarks Anthropic chose, several in a single run per model. CyScenarioBench used different tasks per lab, so the Mythos-versus-GPT-6 Cyber line is loose. The tester quotes are vendor-selected. Gemini 4 Carbon is still in limited release, so its published scores arrive without independent replication. And the rate card understates the bill: caching mix, retries, and effort settings move spend more than the list price does.

Key Takeaways

• Claude Fable 5.2 and Mythos 5.2, the same model with two safeguard profiles, launched Nov 9 at $8/$40 per million tokens, 20% below the 5.1 pair, with outputs up to 30% faster and an optional 2.5x Fast mode at $16/$80.
• FrontierCode 2.0: 49.1% at Xhigh, ahead of Opus 5.5 (45.5%), GPT-6 Astra (42.8%), and Gemini 4 Carbon (41.1%). Cost per task ran $2.48 on Medium, $6.72 on High, $14.19 on Xhigh, and $44.70 on Max, where the score dropped to 45.7%.
• Two losses matter: Carbon wins AutomationBench 53.9% to 51.4%, and Opus 5.5 wins AA-Briefcase head-to-head Elo 1,800 to 1,785 despite losing the rubric 61.2% to 66.1%.
• Anthropic says it capped the model at Opus 5.5 level on kernel optimization and LLM architecture research, the capabilities that most directly accelerate AI development, while leaving coding, hardware, and ordinary ML research intact.
• Scale demos: about 1,000 agents formalized Fermat's Last Theorem in 30 hours where an earlier research model needed 11 days; an empty-folder prompt produced a booting OS after eight days; over 90% of merged changes in Anthropic's core codebase are written by Claude.
• Sandbox-escape credential use fell to 3.1% of scenarios from 21% for Mythos 5.1, and false-positive blocks fell about 55%. All from Anthropic's own evaluations, pending outside replication.

Sources: Anthropic launch announcement, Claude Fable 5.2 and Mythos 5.2 system card, VentureBeat, The New Stack