Opus 5.5 takes first place on the Vals Index

Vals AI's leaderboard put Anthropic's Claude Opus 5.5 first on the Vals Index, but by a narrow margin. The table attached to Vals AI's post, dated 22 September 2026, gives Opus 5.5 69.69% accuracy, Claude Fable 5.1 68.83% and Claude Opus 5 67.21%, so Anthropic holds the top three places. The two-point gain Vals cites over Opus 5 is about two and a half percentage points. OpenAI's GPT-6 Astra sits fourth at 66.61%. Opus 5.5 also costs more to run than Opus 5 in the same table, $22.30 per test against $18.81.

Vals describes the index as a combination of its evaluations of complex work in finance, software engineering, law and tax, meant to measure AI's potential economic impact, in a later blog post. The index has also changed since this ranking. A week later Vals released version 2.1, which replaced Terminal Bench 2.1 with Terminal Bench 4 and added its tax agent benchmark as a separate category, so Opus 5.5's first place was earned on the earlier mix.

The ranking was provisional when posted: Vals noted that GPT-6 Sol results were still to come. By mid-October, in the same Vals blog post, the revised index placed Google's Gemini 4 Argon first, Claude Sonnet 5.5 second and Opus 5.5 third. A sub-point lead on a composite score is a reason to include Opus 5.5 in a shortlist for this kind of work, not proof that it beats the models just behind it on any particular task.