| 1. Claude Fable 5 (High) |
| 2. GPT 5.6 Sol (xHigh) |
| 3. Claude Opus 4.8 (Thinking) |
| 4. GPT 5.5 (xHigh) |
| 5. Claude Sonnet 5 (High) |
The BriefMoonshot AI, a Chinese lab, released Kimi K3 yesterday, and independent testers put it just behind the very best closed models from Anthropic and OpenAI while it costs a fraction as much to run. It is a 2.8-trillion-parameter mixture-of-experts model (only about 50 billion parameters fire on any single request, so it runs far leaner than its size suggests), it handles text, images, and video, and it holds a million tokens of context at once. The API is live today; the full open weights, the part that lets a company download the model and run it on its own hardware, are promised by July 27. It matters because a frontier-class model you can eventually host yourself, at China-scale pricing, changes the math for anyone weighing whether to build on a rented API or bring AI in-house.
Level UpWhen a new model lands, the launch benchmarks are marketing. The durable skill is having your own small, repeatable test that you run against every model, so you judge each one on work that actually resembles yours. Simon Willison has run the same quirky prompt (generate an SVG of a pelican riding a bicycle) against every release for two years, and at a glance it tells him more than any leaderboard. Pick one task from your real work this week, write it down as a fixed prompt, and run it against two or three models. In twenty minutes you will trust your own eyes more than any launch chart. Try it: Simon Willison's pelican test, applied to Kimi K3