Claude Opus 5 commercial tested
Second on the extended CymraegBench leaderboard with an overall score of 85.16.
The facts
| Developer | Anthropic |
|---|---|
| Released | 24 July 2026 |
| API ID | claude-opus-5 |
| Price | $5 / $25 per million tokens (input/output) |
| Context | 1 million tokens; up to 128,000 output tokens |
| Licence | Proprietary (API) |
Model sources: Claude Platform release notes and the model overview.
CymraegBench result
| Overall score | Welsh reasoning | Practical Welsh | Translation BLEU |
|---|---|---|---|
| 85.16 | 89.77 | 80.56 | 70.37 |
The same sample was used as for the other leading models, selected deterministically to return the same cases each time: up to 100 cases from each of the 19 core suites, with random seed 1. The settings were effort=low explicitly, as in the Fable 5 test, and a minimum output allowance of 512 tokens was used because adaptive reasoning is enabled by default. The 232 CCC cases do not contribute to the overall score.
Opus 5 is second on the extended leaderboard: 4.97 points behind GPT-6 Astra, 2.15 points ahead of Muse Spark 1.2, and 2.22 ahead of GPT-5.6 Sol. It completed all 2,053 cases without refusal or invalid output. The API returned one transient empty response; a retry on the same case succeeded when resuming from saved progress.
CCC Language Audit
| Subset | Score |
|---|---|
| Correction and explanation | 91.48 |
| Idiom | 96.88 |
| Terminology | 96.88 |
| Summarisation and quotation | 87.71 |
| References | 72.34 |
| Welsh prompt | 100.00 |
| English prompt | 100.00 |
| Average across the seven subsets | 92.18 |
Corrections were strong (96.88), but the explanation score was 60.94. On the references test, Opus 5 did not invent a source ID, but its answer-content score was 64.06. The profile shows why the audit keeps the dimensions separate.
See the full leaderboard, CCC Language Audit and the open results.