Intelligence Index v5 widens closed-open gap to 13 points

Intelligence Index v5 marks a structural change in how the composite is assembled, not just a reshuffling of scores. By building most of the Index on private datasets, Artificial Analysis reduces the chance that model makers tune against the exact questions being asked, which matters for anyone treating the ranking as an independent check rather than a marketing number. The two additions point in different directions: FrontierCode 2.0, contributed by Cognition, deliberately gives models vaguely specified prompts so that judgment, not just spec-following, is measured, while Terminal-Bench-Science puts 70 agentic research tasks in terminal environments.

The score movements cut both ways. OpenAI's models are rated higher under v5, while some models that ranked well in v4.3, including Meta's Muse Wonder Full(max) and Kimi K3 Pro, have slipped. On the new benchmark results, Claude Opus 5.5 (max) leads FrontierCode 2.0 with GPT-6 Astra (max) second and Gemini 4 Carbon sixth, whereas Gemini 4 Carbon tops Terminal-Bench-Science.

The headline consequence is the closed-versus-open spread, which grew from 3 points to 13. The top open model at the time of writing is Kimi K3.1 Pro at 46, well below Gemini 4 Carbon's 59, and that model has not yet been publicly released; it is expected to be published free of charge soon, so the open-side figure may shift once it is broadly available.