Claude Fable 5.1 holds both crowns again, best available model as well as best measured model, taking them back from GPT-6 two days after losing them…
The Aggregate Digest — Sunday, September 6, 2026
Claude Fable 5.1 holds both crowns again, best available model as well as best measured model, taking them back from GPT-6 two days after losing them, 1815 to 1805. The results that moved it today are agentic. It is first of 59 on Agent Arena, where real tool-use sessions are scored on the net improvement the agent delivered, first on its sub-boards for sessions that ended in success and for praise against complaints, at 22.45 and 42.45 percent, and first of 52 on Multi-turn Debate (Lechmazur), which has models argue both sides of policy propositions across more than 1,000 judged debates, ahead of Claude Fable 5 and Claude Opus 5.
GPT-6 arrived on Friday at 1815 and took both crowns by nine points. It was rated 1805 on Saturday, when Claude Fable 5.1 was already three points ahead, a gap too small to move a crown, which takes five. Today the gap is ten. The two have traded the crowns twice in four days. GPT-6 keeps the boards it took on arrival, FrontierMath Tier 4 (v2) at 97.56 percent to Claude Fable 5.1's 87.8, ARC-AGI-3 at 99.95 and GPQA Diamond in Artificial Analysis' testing at 96.26, and it added first of 126 on Chatbot Arena (Code) today, plus first on WebDev Arena's React and Frontend boards, where developers vote on two models' builds of the same app side by side. Claude Fable 5.1 sits second on both. GPT-6 also leads the HTML board there. On UGI's Willingness board, which scores how readily a model engages with sensitive topics, GPT-6 is 997th of 1,309.
Claude Fable 5.1 has weak boards of its own. It is 26th of 59 on Agent Arena's Steerability sub-board, how well the agent takes user redirection mid-session, and 14th of 27 on HieroglyphBench, which asks a model to transcribe photographed columns of Egyptian hieroglyphs. Claude Fable 5.1 gets 20.4 percent of the signs right there, where GPT-6 gets 40 and Gemini 3.7 Flash leads at 55. Kimi K3 has been the best open-weights model since July 27 and is rated 1757.
Since last time. Gemini 3.8 Flash arrived on Thursday and is third among available models at 1784, one point above GPT-5.5 Pro. It is first of 1,252 on UGI's Writing subscore and first of 57 on Vals AI's Finance Agent v2, which has an agent research SEC filings the way an analyst would. Muse Spark 1.3 is tied for seventh with Gemini 3.7 Flash at 1777 and leads Tau3 Banking, 203 entries of policy-bound banking support, and ComplexConstraints, 72 entries on enterprise instructions whose constraints depend on one another. The week's one new benchmark is the Last Translation Benchmark, 3,456 human-written cases across at least 58 language pairs built to break machine translation, each checked by a hand-written rule. Its human reference clears 99.9 percent. Gemini 3.1 Pro (Preview) leads the 25 entries at 39.3, with GPT-5.6 Sol next at 28.1.
