Claude Fable 5.2 is a small, real upgrade over Fable 5.1: in our test it won 4 of 8 prompts against its predecessor and lost none, and it won 2 of 8 against GPT-6 Astra with no losses. It also costs less per token, $8 input and $40 output per million tokens against $10 and $50 for both rivals. The catch is that it writes longer answers, so a typical answer cost us slightly more than 5.1's.
| Fable 5.2 | Fable 5.1 | GPT-6 Astra | |
|---|---|---|---|
| Input / output per 1M tokens | $8 / $40 | $10 / $50 | $10 / $50 |
| Prompts won by Fable 5.2 (of 8) | n/a | 4 | 2 |
| Prompts won by the rival (of 8) | n/a | 0 | 0 |
| Avg cost per answer | $0.0258 | $0.0215 | $0.0143 |
| Avg latency | 7.8 s | 6.4 s | 7.6 s |
| Avg output tokens | 613 | 398 | 265 |
Prices come from OpenRouter's model list and match Anthropic's announcement. The other figures are from our runs below.
What Anthropic shipped
Anthropic released Fable 5.2 and Mythos 5.2 on 2026-11-09. It describes them as the same underlying model: Fable is open to everyone, with cybersecurity and biology safeguards, while Mythos goes to vetted organizations. The API ID is claude-fable-5-2. Anthropic's announcement also lists a fast mode at $16/$80 per million tokens, a 20% list-price cut from 5.1, and about 55% fewer safeguard blocks than Fable 5.1.
Anthropic reports 49.1% on FrontierCode 2.0 at Xhigh effort against 45.5% for Opus 5.5. Those are vendor numbers, and we did not test them. Our test covers everyday work, not frontier research.
How we tested
We wrote 8 prompts a small team would actually send: a Postgres job-queue race condition, a TypeScript compile error, a refund email under a strict policy, a window-function query, a constrained summary, an inventory puzzle, a date regex, and a JSON ticket triage. We ran Fable 5.2 head-to-head against anthropic/claude-fable-5.1 and openai/gpt-6-astra through OpenRouter, all with a 2,000-token cap. The judge was anthropic/claude-sonnet-5.5, which judged every prompt twice with the answer positions swapped. A win counts only when both orders agree; otherwise it is a tie. Total spend for the post was $1.10 of a $5 cap, on 2026-11-09. Eight prompts is a small sample, so read the results as a direction, not a ranking. The saved runs are in the repo under server/blog/runs/claude-fable-5-2-review-2026/.
Fable 5.2 vs Fable 5.1
Fable 5.2 won 4, 5.1 won none, and 4 prompts ended as ties or judge disagreements. The wins came on the TypeScript fix, the refund email, the SQL query and the ticket triage. On the SQL prompt the judge said:
Response A adds o.id ASC as a final tiebreaker. The request says ties are broken by earliest created_at, so if two orders share both total_cents and created_at, Response B's ordering is nondeterministic. A makes the result deterministic in that case, which is a small correctness advantage.
That is the pattern: the same core answer, with a few more edge cases covered. Fable 5.2 wrote about 54% more output tokens (613 against 398), so its average answer cost $0.0258 against $0.0215 even with the lower price. If your prompts reward careful, complete answers, you get them. If you pay per token for short tasks, set a lower effort or a tighter output limit.
For context on how the family has moved, see our Claude Fable 5 review.
Fable 5.2 vs GPT-6 Astra
This was closer. Six of eight prompts ended as ties, and Fable 5.2 took the other two: the regex and the ticket triage. On the regex prompt the judge caught a factual slip in the rival answer:
Response B's regex works, but its justification is false. It says JavaScript's
$can allow a trailing newline, which is true in Perl and Python but not in JavaScript. [...] Because B states an incorrect fact about JavaScript and A does not, A is better on correctness.
On triage, the judge preferred "other" over "bug" for a leaked API key, and called that a slight edge. Astra answered in 265 tokens on average against Fable 5.2's 650 and cost $0.0143 per answer against $0.0273. So Astra is the cheaper default when ties are common, and Fable 5.2 is the one to pick when a wrong detail costs you.
Subscription vs API
Anthropic's pricing page shows Fable 5.2 is not in the Free plan, and Pro ($20 monthly, or $17 billed annually) reaches it only through usage credits. Max ($100 a month) and Ultra ($200 a month) include it, up to half of the weekly limit, plus $100 and $200 of API credits respectively. At our average of $0.0258 per answer, $100 of API credits covers roughly 3,900 answers like ours, so heavy users break even on Max quickly.
Verdict
Switch from Fable 5.1 to 5.2 if you were already paying for Fable: it won 4 prompts, lost none, and the per-token price dropped 20%. Skip it as a default for cheap, short tasks, where Astra's shorter answers cost about half as much per response in our run. For the cheapest option in the lineup, see our roundup of the best free LLMs in 2026. We have not tested Mythos 5.2, which is limited to verified organizations.