Open weights
An open-weight model finally took the crown
Buried under the frontier-launch noise was the story I care most about: on September 21, Xiaomi released MiMo-V2.6 Pro under an MIT license — 1.02T total parameters, 42B active, 1M context — and it scored 46 on the Artificial Analysis Intelligence Index, the highest ever for an open-weight model, ahead of Kimi K3 and Qwen3.8 Max. At ~$0.13 per task, it's also embarrassingly cheap. Details in the September dispatch.
Why this is the milestone, not the GPT-6 generation
Closed-frontier launches happen every quarter; the gap between "best closed" and "best open" is the number that determines who gets to build what without permission. That gap just effectively closed to a rounding error for most real workloads. Three consequences:
- Sovereignty stops being expensive. Teams that couldn't ship data to an API — healthcare, defense, legal — now have a self-hostable model within spitting distance of frontier quality.
- The price floor is now structural. You can always fall back to $0 marginal cost. Every vendor pricing decision has to compete with free, forever.
- The moat question inverts. If weights are a commodity, the differentiators are exactly the unglamorous ones: your data, your evals, your tool integrations.
One caveat
Index scores compress real differences. A 3-point gap on a benchmark says nothing about the failure modes you'll hit in production — and open models still lag on the newest modalities. Run your own evals before migrating anything. But the direction of travel is unmistakable: 2025's gap was a chasm; September 2026's is a crack.