Cheap, Capable, Lightly Governed

Frontier AI coding is now a commodity input, and the safety paperwork is not. The shape this quarter is cheap, capable, lightly governed.

Per the Last Week in AI show notes, Chinese open-source models crossed thirty percent of weekly OpenRouter tokens (a routing service many developers use to pick models) as cost pressure rose. In the same week, SpaceX AI's Grok 4.5 launched as an Opus-class coder (a benchmark tier above Anthropic's Claude Opus) at a fraction of prior Opus-class pricing, and Meta's Muse Spark 1.1 cut aggressive pricing with large coding and cyber benchmark gains. Three labs, one direction, one week.

The more useful read — inferred from the convergence of three labs in one week — is that when frontier capability becomes cheap enough to be a commodity input, governance stops being a per-lab problem and becomes a coordination problem. The safety evaluation moves with the model. The public oversight, the policy framing, the legal liability: those move on a different clock. The gap between model speed and oversight speed is the new product surface.

Builders and buyers get a thirty-day diagnostic. Name the model behind any AI product you ship or sign for, read its safety evaluation if one exists, and price it at scale. The winners of this regime are the platforms that route. The losers are the oversight regimes still built for one lab, one country, one release.