Opus 5 tops the smart list with a weak edge, and single job costs are 26% below Fable 5

On July 25, the third-party evaluation agency Artificial Analysis published Claude Opus 5 achievements. It scored 61 points in the smart index that combined the nine tests, with a narrow lead of 60 points in Fable 5. GPT-5.6 Sol score 59 and Kimi K3 score 57. The average cost per mission for Opus 5 is $2.03, 26 per cent below $2.75 for Fable 5. It tops up in the assessment of the knowledge work of the GDP val-AAA v2 and AA-Briefcase, combined with Claude Code, and lists the first of the programming Agent indices. Terminal-Bench v2.1 score 89%, roughly levelling GPT-5.6 Sol. The model provides a 5-storey reasoning strength. From low to max, output token is about eight times different, and GDP val-AAA v2 is 407 Elo. Users can trade more Token for more performance, and can press low costs. So is the slab. Opus 5 is still lagging behind Fable 5. In the AA-Omniscence test, its hallucination rate rose to 50 per cent, 14 percentage points higher than Opus 4.8. The value-for-money ratio of the lower reasoning slots is also slightly lower than the GPT-5.6 series。