Claude Fable 5.2 Costs 2x Opus 5.5. When It's Worth It

Anthropic's Claude Fable 5.2 lists at $8 per million input tokens and $40 output — twice Opus 5.5 — and a plain model-ID swap will start returning 400 errors.

It shipped on 9 November 2026 alongside Claude Mythos 5.2, a second configuration of the same model, and is available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and GitHub Copilot. The buying question is narrower than the launch post suggests.

What the price buys

Anthropic reports outputs arriving up to 30% faster than the 5.1 generation, with an optional fast mode at $16 and $80 that runs up to 2.5 times faster. The published spec sheet puts the context window at 1 million tokens with 128,000 max output, a June 2026 knowledge cutoff, and a default effort of high. Batch pricing is 50% cheaper. Zero data retention needs explicit approval — the default is a required 30 days.

Model	Input / output per 1M	Default effort
Claude Fable 5.2	$8 / $40	high
Claude Opus 5.5	$4 / $20	medium
Claude Sonnet 5.5	$2 / $10	high

On benchmarks, the vendor reports Fable 5.2 at 49.1% on FrontierCode 2.0 at Xhigh effort and 69.9% on FrontierSWE v2, against 62.3% for Opus 5.5 on that second test. Read the table with its own footnote attached: all scores show each model's highest-available effort unless stated otherwise, which is why the headline numbers sit close together.

The effort tier is the actual pricing decision

The accuracy-against-cost charts answer a different question than the benchmark table, and it is the question most teams face. At High effort, Fable 5.2 scores 45.3% on FrontierCode 2.0 against 45.5% for Opus 5.5 at Max — roughly half the cost per task. On Humanity's Last Exam, Fable at Low effort matches Opus 5.5's top score for less than one-tenth of the cost per task; on WANDR at Medium, about 40%.

Medium is the default in the Claude apps and in Claude Code. So the same model sells at five effort levels with the default in the middle, and between Low and High the cost curve is steeper than the accuracy curve. That makes the tier a per-workload budget setting, not a global switch. The measurement that matters is cost per accepted change: a pipeline burning Xhigh effort on work a reviewer rejects twice has spent more than one running High on the same task.

Where it loses

Gemini 4 Carbon leads AutomationBench at 53.9% against Fable 5.2's 51.4%. And Anthropic's own framing is that Fable and Mythos suit the most ambitious, open-ended technical work, while most coding and knowledge tasks should start on Opus 5.5 and move up or down from there. That is unusually direct guidance against the more expensive SKU, and it should anchor your default.

The migration will 400 your backend

Fable 5.2 no longer accepts sampling parameters, forced tool calls, or disabled thinking. Concretely, per the migration write-up: temperature, top_p and top_k return 400; tool_choice: {type: "any"} returns 400, so use auto and name the tool in the prompt; thinking: {type: "disabled"} and budget_tokens return 400, with depth controlled by effort instead. Assistant prefill also 400s. A small max_tokens can stop the response after thinking with no text at all.

Two more worth catching before they page someone. Refusals now come back as an ordinary response with stop_reason: "refusal", so a backend that only handles exceptions will treat a refusal as success. And thinking blocks are bound to the prior conversation history: change the system prompt, the tool list, or older messages between requests and the API rejects the old blocks with a signature error — a check that now applies to all accounts.

What to do this week

Grep for the 400-triggering parameters before you touch a model ID. Build profiles rather than a global swap: Opus 5.5 for everyday agent steps and reviews, Fable for large refactors and long runs where Opus fails at high effort. Then run your own acceptance test — every figure above is vendor-reported at the vendor's chosen effort, and the only number that settles a 2x premium is your cost per accepted change.