Kimi K3, the best open-weights model, ties for first on xDailyBench's everyday consultation tasks, ahead of GPT-5.5 and GPT-5.6 Sol
The Aggregate Digest — Sunday, September 27, 2026
Kimi K3, the best open-weights model, ties for first on xDailyBench's everyday consultation tasks, ahead of GPT-5.5 and GPT-5.6 Sol. xDailyBench scores professional consultation on real daily problems, each task against its own rubric. Kimi K3 and Seed-2.1-Evolving both score 75.6 percent over the 11 entries, then GPT-5.5 at 72.6, GPT-5.6 Sol at 72.4 and Claude Opus 4.7 at 70.8. The board also splits each task's requirements into the ones the user stated and the ones left unsaid, and the unsaid half is where every model drops: Kimi K3 leads it at 67.6 percent against its 78.2 on the stated half. Claude Opus 5.5, GPT-6 and Claude Fable 5.1 were not tested. Kimi K3 is #24 on the table at 1899.
Planted web pages still end up in deep-search answers about a third of the time. HAE-GEO runs deep-search agents against web evidence poisoned at three levels and counts how often the planted evidence is used in the answer. At the mildest level the least fooled of ten entries is Qwen 3.8 27B at 26.6 percent, then Claude Opus 4.8 at 29.7 and GPT-5.6 Sol at 33.3. The most aggressive level fools Qwen 3.8 27B 31.7 percent of the time and Claude Opus 4.8 51.6. Answering well from clean evidence is no protection: Qwen 3.6 35B A3B ties GPT-5.6 Sol for the best clean-answer score and is the most fooled entry at every level, 59.7 percent at the mildest and 69.4 at the most aggressive.
Claude Opus 5.5's crown now has arena votes behind it, and GPT-6 keeps the image boards. Claude Opus 5.5 took best available model on Friday and holds it at 2052 to GPT-6's 2039. This week its max-effort run took eight of WebDev Arena's eleven category boards and went first of 134 on Chatbot Arena's code board, ahead of GPT-6 in both. Today the high-effort run is first of 409 on Chatbot Arena's text board, ahead of Claude Opus 4.6 and Claude Fable 5. Where the task is an image, GPT-6 stays ahead. ActiveVision makes a model scan and transform images over several steps, and GPT-6 scores 76.5 percent to 41.2. Claude Mythos 5 is still the best measured model, at 2062.
Elsewhere. SciTrue checks scientific claims against the paper they cite, and Claude Fable 5 leads all three of its boards, with claim validation at 86.3 percent F1 and attribution at 97.7. Gemma 4 31B, an open model, ties GPT-5.5 and GPT-5.6 Sol on claim validation at 84.5. Search-to-World has an agent search the web for assets and deliver a navigable 3D world, and DeepSeek V4 Flash, #120 on the table, resolves 88.3 percent of requests; GPT-5.6 Terra is next at 72.7 and Claude Fable 5 tenth of 12 at 41.0. The Explicit Edit Benchmark replays 226 exact-edit tasks through coding agents and passes a file only when it matches byte for byte. DeepSeek V4.1 Flash and Muse Spark 1.3 tie for first of 16 at 99.0 percent; GPT-5.6 Sol scores 98.9.
