There is no single AI frontier anymore. There are two — proprietary models behind APIs, and open weights anyone can download — and they move differently. So the Watch now tracks them separately.
October 21, 2026 · 5 min read
Ask "what's the best AI model?" in October 2026 and the honest answer is: which frontier do you mean? The proprietary frontier — OpenAI, Anthropic, Google, xAI — sells capability through APIs. The open-weight frontier — DeepSeek, Z.ai, Moonshot, Xiaomi, Alibaba, Mistral, NVIDIA, and others — gives the weights away and competes on licenses, hardware, and release cadence. They are different games, and a single blended leaderboard hides what matters about both. That is why the Watch now runs two tracks: a Commercial Watch and an Open Source Watch.
On SWE-bench Verified, the classic coding benchmark, the top of the table has effectively converged: Claude Opus 5 posts 96% (BenchLM), GPT-5.6 Sol 96.2% on Vals AI's run, Claude Fable 5 95.0%. The leaderboard operators themselves now flag the benchmark as saturated and contaminated — the field is moving to harder successors like SWE-bench Pro and Terminal-Bench, whose public boards read very differently (Muse Spark 1.1 leads Scale's standardized SWE-bench Pro board at 61.5%). Meanwhile the release drumbeat continues: Google's Gemini 4 Argon went generally available on October 13 at introductory API pricing of $2/$10 per million tokens (Reuters), and Meta launched Muse Wonder on October 12. The commercial questions are now less "who is smartest?" and more "at what price, with what limits, and what happens when it misbehaves?"
The open frontier turns over roughly every six weeks. On the Artificial Analysis Intelligence Index (v4.3.2, read October 5), Xiaomi's MiMo-V2.6-Pro leads all open weights at 46.32 — under a plain MIT license — followed by GLM-5.3 (44.78) and Kimi K3 (43.59). On SWE-bench Verified, DeepSeek's V4-Pro posts 80.6%, the top open score, with MiniMax M3 at 80.5%. But the number that decides whether you can ship is the license: MIT (the DeepSeek V4 line, GLM-5.3-Flash, MiMo-V2.6-Pro) and Apache 2.0 (Qwen3.8-27B, Gemma 4, gpt-oss-120B, Meta's Muse Glimmer 30B) on one side; bespoke, sometimes revenue-gated terms on the GLM, Kimi, and Qwen flagships on the other. October's releases tell the same story: NVIDIA's Nemotron 3.5 family (Oct 15–16) ships open weights sized for a single 24GB GPU, Liquid AI's LFM2.5-24B-A2B (Oct 14) targets agents on laptops, and Mistral Large 4 (Oct 6) is API-only until its weights land at the end of the month.
Because the questions differ. For commercial models you track capability per dollar, rate limits, incidents, and vendor behavior. For open models you track licenses, hardware fit, and who released what this week — the download page matters as much as the benchmark. Mixing them into one ranking answers neither set of questions well. So: two watches, each with dated standings, cited sources, and vendor figures labeled as vendor figures. When a leaderboard rebases its index, the old numbers retire instead of blending in.
The tracks are live: Commercial Watch · Open Source Watch. Receipts are on both pages.