Claude Fable 5.2 by Anthropic
The best model you can buy, a fifth cheaper per token, but still 2.6 times Sonnet per task.
The pain: Some jobs beat even Opus: the long agent run that ends half-done, the hard bug it circles. Until now the only model above it cost $10 in and $50 out per million tokens.
The fix: Fable 5.2 is first on the Artificial Analysis index at 65, at $8 in and $40 out, 20% cheaper than Fable 5.1, and on Bedrock the same day.
Anthropic shipped Claude Fable 5.2 on 9 November, ten weeks after Fable 5.1. It's a point release that matters for one reason: it's now the strongest model you can buy, first on the Artificial Analysis Intelligence Index at 65 at maximum effort, and it's 20% cheaper per token: $8 in and $40 out per million tokens, down from $10 and $50. Mythos 5.2 is the same model with the cyber and biology safeguards relaxed, for vetted organisations only.
That's a price drop at the frontier of intelligence, which is always news here. But read the per-task numbers before you switch anything:
| Model | Index | Cost per task |
|---|---|---|
| Claude Fable 5.2 (max) | 65 | $26.54 |
| Gemini 4 Carbon (high) | 60 | $5.48 |
| Claude Opus 5.5 (max) | 59 | $10.04 |
| GPT-6 Astra (max) | 57 | $6.24 |
| Claude Sonnet 5.5 (max) | 55 | $10.02 |
Fable at max effort buys ten points over Sonnet for 2.6 times the cost per task. At high or xhigh effort it scores 62, but the index doesn't publish a cost for those. Anthropic's own guidance is the sensible one: start with Opus, and move to Fable only when Opus at high effort fails your evals. Opus costs the same per task as Sonnet and scores four points higher, which is the quieter story of the season.
On Bedrock it arrived the same day, on the global and US profiles. Read the data terms before you route client work there: Bedrock requires the aws_review retention mode for this model, AWS keeps all Fable traffic for 30 days, and flagged traffic can be read by AWS staff. Anthropic's own API treats it as a covered model with 30-day retention too, so zero data retention isn't on offer anywhere.
Gotchas, with fixes:
- Rate limits are a quarter of Sonnet's at the Start tier: 500K input and 100K output tokens a minute. Cache your prompts, since cache reads don't count, and ramp traffic up gradually.
- Breaking API changes:
tool_choiceset toanyor a named tool returns a 400, so do disabled thinking and abudget_tokenssetting, and temperature and top-p are gone. Useautowith strict tools, drop the sampling parameters, and keep the conversation history append-only. - Effort defaults to high, not medium. Set it per request, and keep max for the jobs that justify $26.
- The tokenizer (since Opus 4.7) produces about 30% more tokens for the same text than older Claude models. Re-measure your own prompts before you compare bills.
What it can do now: work unattended for hours with subagents (Anthropic's claim), a fast mode at 2.5 times the output speed for double the price ($16 / $80), thinking preserved across turns, and an advisor tool with a Fable advisor (beta). It's guaranteed until at least 9 November 2027.
My verdict: watching. This is the model for the job Opus can't finish, not a new default. For most production work, Opus or Sonnet costs less than half as much per task and does the job.