Claude Opus 5

On the Artificial Analysis Intelligence Index, as reported in Mozilla's The state of open source AI (v1.0.1, July 2026), the best closed model scores 61, and that model is Claude Opus 5; the best open one, Kimi K3, scores 57 and ranks fourth overall. A blind human evaluation of code by Surge AI, cited in Mistral AI's launch post for Mistral Large 4, puts Claude Opus 5 first of five at 4.22, ahead of Mistral Large 4 at 3.74.

The same model supplies the sharpest cost warning in the Block benchmarks published by Atish Patel on 6 August 2026. Run solo on Terminal-Bench 2.1 at xhigh effort, Claude Opus 5 cost $140.63, the most expensive run of the panel, for 75.0%, which placed it below six cheaper runs. Over-reasoning had timed it out on 17 of the 88 tasks. GPT-5.6 Luna reached the same 75.0% at $4.98. Patel keeps Claude Opus 5 in his SmartBee tier, the one that handles overview, arbitration and escalation, but recommends it at medium effort.

Anthropic's own figures place it behind its successor. On Terminal-Bench-Science 0.1, Claude Opus 5 scores 29.0% where Claude Fable 5.1, announced 1 September 2026, scores 52.6%. What remains of its stated role is narrower: fallback for cyber and biology tasks that have been redirected. A model at the top of one index and sixth on price-performance in another is two different purchases.