The Aggregate Digest — Tuesday, September 15, 2026

Qwen 3.8 Max (0902) is the ninth-best available model one day after appearing on the site, and building web apps is where it comes closest to the leaders. On arena.ai's WebDev Arena, where two models build the same web app and human voters pick the rendered result they prefer, it is second of 127 on data and analytics apps behind GPT-6 at max effort. It is fourth on the frontend and React boards and fourth again on gaming and simulations, boards of 112 to 128 entries. On Chatbot Arena (Code) it is fourth of 128, behind GPT-6, Claude Opus 5 and Claude Fable 5.1, each at max effort. Its rating is 2024 on 36 results, 49 points above the Qwen 3.8 Max already on the table, which is tied for 18th at 1975 on 430 results and trails it on every one of those arena boards.

Away from the arenas the record is thin, the thinnest in the top 20. It scores 100 percent on Epoch AI's OTIS Mock AIME 2024-25, olympiad-level mock competition problems, a perfect score shared with at least six other entries of 290. On MLS-Bench Lite, 30 tasks on inventing machine-learning methods that generalise and scale, it is first of 14 at 50.1 to Claude Fable 5's 49.9. Its 66.3 on SWE Atlas leads that 38-entry board but is self-reported. On AI Coding Daily's OpenCode board it is third of 20 with 48.7 of 60 points. Kimi K3 at high effort leads that board with 50.85 and GLM-5.3 is second.

GPT-5.6 Pro Sol, on the site since August 1, is eighth at 2026 on 55 results, two points ahead of it. It is first of 55 on AIMultiple's FinanceReasoning (Hard), 238 multi-step quantitative finance questions, at 90.76 percent, with Claude Fable 5, Claude Fable 5.1 and GPT-5.6 Sol tied next at 90.34. On PrinzBench, 33 questions on legal research and locating obscure public information, scored out of 99, it is first of 29 with 91 to GPT-5.5 Pro's 82. GPT-6 remains the best available model at 2123, with Claude Fable 5.1 eight points behind at 2115.

Read the full digest on The Aggregate →