Claude Opus 5: The Smartest Model According to Artificial Analysis – What's Behind the Scores

Anthropic unveiled Claude Opus 5 on July 24, 2026, positioned between Sonnet 5 and the restricted Fable/Mythos 5. It features a 1M token context window, 128K output, and reasoning enabled by default with five effort levels: low, medium, high, extra, max. On internal benchmarks, Opus 5 scores 43.3% on Frontier-Bench v0.1 (agentic coding) versus 18.7% for Opus 4.8 and 33.7% for Fable 5; on ARC-AGI-3 (abstract reasoning) it achieves 30.2% compared to GPT-5.6 Sol's 7.8%. However, independent Artificial Analysis Intelligence Index gives Opus 5 a score of 61 (max effort), just one point above Fable 5 (60) and two above GPT-5.6 Sol (59), declaring it the smartest model by a narrow margin. The model is described as noticeably slow and verbose: time to first token on max effort exceeds one minute, and total token consumption for the benchmark suite was ~100M vs ~63M median. Despite wins, Opus 5 loses on DeepSWE v1.1 agentic coding (68.8% vs GPT-5.6 Sol's 72.7%), HealthBench Professional (59.8% vs Mythos 5's 66.0%), and Legal Agent Benchmark (11.7% vs Fable 5's 13.3%). Anthropic's report highlights a case where Opus 5, given a blueprint without vision, wrote its own computer vision program from raw pixel data to solve the task, while competitors failed. On the negative side, in a protein design test with a $10K budget and 24 hours, Opus 5 delivered nothing in one run and only 17 unranked variants in another, falling victim to unproductive self-verification loops. The company warns that max effort may suffer from overthinking, and recommends starting at extra for coding and high for other tasks.