Artificial Analysis releases Intelligence Index v5 as open-source model gap widens to 12 points

AI benchmark firm Artificial Analysis has released Intelligence Index v5, its biggest overhaul since v4 launched. For the first time, 60% of the test weight relies on private test sets or hidden answers, to stop models from being trained on public test questions. In coding, the new FrontierCode 2.0 uses 132 tasks from real open-source repos. Code changes are scored by repo maintainers on whether they would merge them, not just whether tests pass. DeepSeek V4.1 Flash scored 28%, the highest among models with released weights. On the overall score, DeepSeek V4.1 Flash and GLM-5.3 both got 32 points, behind leader Claude Opus 5.5 at 58. The gap between the best open-source model and the top model grew from 3 points to 12. Effort levels also matter more now: Claude Opus 5.5 costs about 10 times more at max effort than at low effort.