No model router reaches the top ten of Vals AI's new test at default settings
The Aggregate Digest — Sunday, October 25, 2026
No model router reaches the top ten of Vals AI's new test at default settings
Vals AI tested five routing systems, which choose among a pool of models, on eight tasks from its Vals Index. The best of them places 13th of 27, so on these tasks picking one strong model scored higher than handing the choice to a router.
The best router is Switchyard, at 57.8%. Gemini 4 Argon leads the board at 69.2% and Claude Sonnet 5.5 is second at 69.0%. Claude Haiku 5.5 is tenth at 58.7%. The other routers score from 48.1% to 56.5%, with openrouter/auto at 52.2%. Each router ran with its default model pool and its default balance of quality against cost, and a refusal counted as a failure with no backup, unlike some of Vals AI's standard index runs.
Vals AI Router Index v1.0, October 25, 2026. Eight Vals Index v2.2 tasks with the same weights and limits. Routers in teal at their default settings; selected single models in grey, with no setting listed. 27 entries; Claude Haiku 5.5 is tenth.
Gemini 4 Argon also leads Kaggle's game ratings and a medical image exam. Kaggle Game Arena pools each model's matches across its games into one rating. The games include text chess, Four in a Row, heads-up poker and Werewolf. Gemini 4 Argon is first on that rating, ahead of GPT-6 Astra. On the separate game boards it is first at Four in a Row and fourth at chess. It has no poker or Werewolf result. Natomy runs MedXpertQA's 2,000 expert-level medical exam questions that come with clinical images and five answer choices. Gemini 4 Argon answers 90.5% correctly, against 87.4% for GPT-6 Astra. It is #8 on the table, with an Aggregate Rating of 1963.
Natomy's MedXpertQA multimodal run, October 25, 2026. 2,000 medical exam questions with clinical images, five answer choices, zero-shot with chain-of-thought. Claude Opus 5.5 at high effort; the other rows list no setting. 14 entries.
Muse Wonder Full leads Vals AI's new legal index. The index combines Vals AI's legal research, legal agent and tax agent tests, weighted by 2022 US law-office revenue. Muse Wonder Full at max effort scores 77.86% and Gemini 4 Argon at high effort 76.72%. Claude Opus 5.5 is seventh at 72.34%, also at max effort. Artificial Analysis also runs Harvey's legal agent tasks and counts a task only when it passes every criterion and a hallucination check. Muse Wonder Full at max effort does that on 13.06% of tasks, the highest of 28 entries, and Gemini 4 Argon at high effort on 3.89%.
Elsewhere. MiniMax-M3.1-Flash joined the table this week and is now #39. Artificial Analysis' AutomationBench scores agents on workplace objectives, with deductions for guardrail violations. MiniMax-M3.1-Flash is third there at 72.52%, after Gemini 4 Argon at high effort with 77.51% and Kimi K3.1 Pro with 73.3%. Artificial Analysis puts its cost at $0.22 per Intelligence Index task, against $1.99 for Gemini 4 Argon at high effort and $2.11 for Kimi K3.1 Pro.
