Simon Willison's six-month LLM recap: coding agents work, open weights catch up

Simon Willison's five-minute lightning talk at PyCon US 2026 is the kind of summary worth reading because the half-year it covers contained two real shifts. The first is that coding agents stopped being a sometimes-tool. They moved from "often-work to mostly-work," the threshold where a developer can keep one running through the day without spending most of the time cleaning up after it. The second is that open-weight models stopped lagging by an obvious gap. Willison points at Qwen3.6-35B-A3B, a 20.9GB model that runs on his laptop, and notes that it drew him a better pelican-on-a-bicycle than Claude Opus 4.7 did, the same informal test he has been using for years to compare models.

The talk also captures how volatile the frontier got. Between November 2025 and the spring, the title of "best model" changed hands five times across Anthropic, OpenAI, and Google. Willison reads that as the result of new work in reinforcement learning from verifiable rewards starting to land in production. He also gives a nod to the cultural moment around OpenClaw, the personal AI assistant whose name has been borrowed to refer to the whole category in casual conversation.

What makes the talk durable is that it does not try to predict the next six months. It just names what changed, with one tested example per claim.

Why it matters

If you have been waiting to evaluate a local open-weight model against your usual hosted choice, Willison's pelican result is a cheap nudge to retest now: a model that runs on a single laptop GPU is a meaningfully different procurement story from a 2025 open weight.