GPT-6 arrived today and is now the best available model and the best measured model, taking both crowns from Claude Fable 5.1 at 1815 to 1806
The Aggregate Digest — Friday, September 4, 2026
GPT-6 arrived today and is now the best available model and the best measured model, taking both crowns from Claude Fable 5.1 at 1815 to 1806. Claude Fable 5.1 had held the best available crown for two days and the best measured crown for one. The nine-point margin rests on 55 results, and those results are the reason to care. GPT-6 is first of 60 on FrontierMath Tier 4 (v2), the hardest unpublished research-level mathematics problems, at 97.56 percent. The previous best there was 87.8. On ARC-AGI-3 an agent is dropped into a game with no instructions and has to work out the rules by playing. GPT-6 is first of 39 there at 99.95 percent. The previous best was Claude Opus 5 at 30.16.
GPT-6 is first of 101 on FrontierMath Tiers 1-3 (v2) at 93.68 percent. Claude Fable 5.1 is second at 90.18. GPT-6 is first of 611 on Artificial Analysis' GPQA Diamond at 96.26 percent, one day after Gemini 3.8 Flash took that board at 95.25. On ARC-AGI-2 only the human panel is above GPT-6, which scores 95 percent and is second of 225. GPT-6 also leads Epoch AI's Chess Puzzles at 72 percent among 220 entries, Epoch AI's ECI among 585, and SimpleQA Verified among 78.
GPT-6 is not first everywhere. It is 51st of 535 on Artificial Analysis' Long Context Reasoning and 33rd of 602 on their SciCode. On their Humanity's Last Exam it is sixth of 602 and Claude Fable 5.1 still leads. Claude Fable 5.1 lost eight points of its own today, 1814 to 1806. The best open-weights model is unchanged: Kimi K3 at 1750.
Muse Spark 1.3 was five results short of a ranking yesterday. Today it has 41 and is ninth among available models at 1766, one point behind GPT-5.6 Sol. Muse Spark 1.3 is first of 203 on Tau3 Banking, where GPT-6 is 22nd. In LLM Stats' testing it is also first on AutomationBench, Job Bench, DeepSWE 1.1 and MRCR v2 (8-needle).
