Claude Fable 5 commercial tested
Eighth on the extended leaderboard, but with an unusual refusal rate on biology questions.
The facts
| Developer | Anthropic |
|---|---|
| Released | 9 June 2026 |
| Class | A ‘Mythos’-class model, positioned by Anthropic above Opus. Fable 5 is the generally available version; Mythos 5 is the restricted-access version with fewer safeguards. |
| Price | $10 / $50 per million tokens (input/output) |
| Licence | Proprietary (API) |
Welsh: no data from Anthropic
The multilingual result quoted in the announcement is around 96% on MGSM, a mathematics benchmark covering ten languages, but not Welsh. Anthropic’s announcement does not mention Welsh, and the Claude family’s 15-language documentation table does not include it either. As with the other models on the leaderboard, this is not evidence of inability; it is evidence of a lack of measurement.
The question is relevant for Fable because its predecessor has already been measured. Opus 4.8 reached 20.0% for valid seven-syllable lines and 13.5% for appropriate+valid+original in Bangor AI’s cynghanedd protocol. Repeating the test with Fable can help measure whether its Welsh ability has changed.
CymraegBench v0.1 result
| Overall score | Welsh reasoning | Practical Welsh | Translation BLEU |
|---|---|---|---|
| 78.64 | 83.73 | 73.56 | 69.46 |
Fable is eighth after the latest results were added, but refused 79 of 1,821 requests (4.34%). Many were school biology questions classified as ‘bio’ requests by Anthropic’s safety system; each refusal scored zero. See the full leaderboard and the open results.
What comes next
The cynghanedd protocol and qualitative tests Coleg Cymraeg report will be the next steps.