Six Months of LLM Progress Distilled Into Five Minutes

The LLM landscape has moved faster in the past six months than any comparable period prior. Willison's condensed recap surfaces the signal from the noise: which model releases actually changed what engineers can build, which tooling shifts matter, and where the frontier moved.

The pattern visible across this period is compression — more capable models at lower cost, smaller context-window constraints, and faster inference. Reasoning models from both Western and Chinese labs landed in production-usable form. Open-weight releases closed the gap with closed APIs in specific task categories. The number of credible coding assistants multiplied.

For engineers making architectural decisions, the implication is practical. Model selection is no longer a one-time choice. A model that was the right call three months ago may no longer be the best fit for latency, cost, or quality on a given task. The evaluation cadence needs to match the release cadence.

For solo founders, the cost curve matters most. Running serious LLM workloads became cheaper. Tasks that were cost-prohibitive at scale in late 2025 are now viable. That changes the unit economics of products built on top of inference APIs.

The tooling layer is stabilizing around a smaller set of patterns: structured output, tool use, and context management. The early chaos of competing abstractions is narrowing. Teams building on top of raw APIs are less disadvantaged relative to those using orchestration frameworks.

Willison's format — a tight oral summary — is itself a signal. The space is moving fast enough that a six-month retrospective fits in five minutes and still captures meaningful shifts. Engineers who have not kept up with releases since late 2025 will find the recap a reliable way to recalibrate without spending a day reading release notes.

The full summary is available at simonwillison.net.

Source