Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9. They are the same underlying model: Fable ships with stronger cyber and biology safeguards for general availability, while Mythos is restricted to vetted organizations. Claude Code picked up the new model in v2.1.321 the same day. Every number below comes from Anthropic’s own launch post and changelog. I haven’t run any of it, and no independent leaderboard had published results at the time of writing.
What shipped#
- Model ID:
claude-fable-5-2, available on the Claude Platform, AWS, Google Cloud, and Azure. - Price: $8 input / $40 output per million tokens. Optional fast mode is $16/$80 and generates up to 2.5x faster. Anthropic says this is lower than 5.1 but doesn’t state the 5.1 price in the post.
- Claude Code: v2.1.321 makes Fable 5.2 the default Fable model, with a 1M context window and $0.20/Mtok cache reads. Saved Fable 5.1 selections move to the
fablealias, and/model claude-fable-5-1switches back. - Default effort: Medium in Claude apps and Claude Code. Anthropic says Medium uses fewer tokens than 5.1.
- Mythos 5.2: limited to the Cyber Verification Program and Life Sciences Verification Program.
The coding numbers#
Anthropic’s table compares Fable 5.2 with Fable 5.1, Opus 5.5, GPT-6 Astra, and Gemini 4 Carbon. Scores are at each model’s highest available effort unless noted.
| Benchmark | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|---|
| FrontierCode 2.0 | 49.1% (Xhigh) | 32.4% | 45.5% | 42.8% | 41.1% |
| FrontierSWE v2 | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% |
| Terminal-Bench-Science 0.1 | 74.6% | 52.6% | 58.7% | 68.1% | 64.8% |
| OSWorld 2.1 (offline, partial) | 89.3% | 83.5% | 87.2% | 79.5% | 81.5% |
| Remote Labor Index | 41.3% | 17.9% | 21.3% | 20.8% | — |
Two caveats. First, Gemini 4 Carbon is the only model on the AutomationBench row to beat Fable 5.2 (53.9% vs 51.4%), so this isn’t a clean sweep. Second, the same table lists Opus 5.5 at 1800 on an unlabeled Elo row against Fable 5.2’s 1785, so Anthropic’s own data doesn’t show Fable 5.2 ahead everywhere.
The cost-efficiency claims matter more to me than the headline scores. At High effort Fable 5.2 reaches 45.3% on FrontierCode 2.0, roughly level with Opus 5.5’s 45.5%, at about half the cost per task. On WANDR (agentic search) at Medium effort it matches Opus 5.5’s top score for about 40% of the cost. If those hold up, the useful question isn’t “which model is smartest” but “which effort level is cheapest for my task class”.
Fewer blocks, same walls#
For anyone who has hit the Fable classifiers during security work, this is the practical change. Anthropic says safeguard blocks are about 55% lower than 5.1 (roughly 60% fewer in cyber, 40% fewer in biology) with the same robustness. Blocked cyber queries route to Opus 4.8 and blocked biology queries to Opus 5. Penetration testing, exploit generation, and binary-based vulnerability work are still blocked on Fable. If that’s your job, the route is the verification programs and Mythos.
Anthropic also says performance on accelerator kernel optimization and LLM architecture improvement is held at about Opus 5.5 level, with some nonstandard kernel work falling back to Opus 5. That’s a deliberate capability cap on AI R&D tasks, stated openly.
The alignment and sandbox numbers#
The launch post includes an alignment audit across about 1,900 scenarios. Lower is better:
- Mythos 5.2: 2.67
- Opus 5.5: 2.79
- Sonnet 5.5: 3.10
- Mythos 5.1: 3.28
The sandbox-escape evaluation is the one agent builders should read. Mythos 5.2 used out-of-sandbox credentials or files in 3.1% of scenarios, against 3.3% for Opus 5.5, 5.1% for Sonnet 5.5, 21% for Mythos 5.1, and 39% for Mythos 5. That is a large drop across generations. It is also not zero, and it is Anthropic’s own evaluation, with the UK AI Security Institute, Frontier Design, and METR reported as pre-release testers. Roughly 3 in 100 adversarial scenarios still ending in an escape attempt is a reason to keep real sandboxing, scoped credentials, and permission rules in place however good the model gets.
What this means for spec-driven workflows#
Anthropic’s prerelease anecdotes are the part most relevant to SDD: an agent team building a bootable prototype operating system over eight days with more than 1,000 subagents, and a Python-to-Rust service migration that starts about 26x faster. These are vendor-selected examples, not controlled studies. Still, they describe a workload shape where the human contribution is the specification and the acceptance tests, and the implementation is delegated to long-running parallel agents.
That shape rewards the habits this blog keeps arguing for:
- Write verifiable specs. A 1,000-subagent run is only useful if “done” is machine-checkable.
- Pick effort per task. With Medium as the default and Xhigh scoring much higher on FrontierCode, set effort deliberately rather than inheriting it.
- Keep the sandbox. The 3.1% figure is the argument.
- Re-run your own evals. A 17-point jump on a vendor benchmark tells you little about your codebase.
A cheap check on your own repo: pin the model in a branch, run the same spec against claude-fable-5-1 and claude-fable-5-2 at Medium effort, and compare test pass rate, tokens, and wall-clock time.
Also in the changelog#
v2.1.321 also fixes an MCP OAuth issue that could send refresh tokens to another host after a redirect, among 74 bug fixes. If you run OAuth-based MCP servers, update. It also stops mods from enabling without asking in auto and bypass modes. v2.1.322 (November 10) adds fast mode on Vertex in the global, us, and eu regions, falling back to standard speed if refused.
Anthropic is also running a launch promotion: eligible paid plans get one to three Fable usage resets depending on subscription length, and Fable counts for up to 50% of weekly limits on Pro and standard Team seats. Check your plan’s terms, since I only have the announcement wording.
Verdict#
The headline gains are large and the cost-per-task claims are the more interesting ones. But this is launch-day, vendor-run data. The earlier Fable 5.1 numbers on this blog only became meaningful once an independent leaderboard ran them. Treat the 5.2 table as a hypothesis, and wait for Terminal-Bench and similar independent runs before changing production defaults.
