What the AI models are actually good at

A transit map of the model landscape. Each coloured line is one thing you might want a model to be good at, and stations along it are the models that can do it, in rank order, best first. White capsules are interchanges: the same model serving three or more different capabilities. Dashed lines in the lower band are the products you actually type into — Claude Code, Cowork, Codex, Cursor, Copilot — and their stations are the models each one can run.

Below the ranked lines, everyone else: the Chinese second wave, Asia-Pacific and European sovereign labs, Western open and research labs, enterprise platforms, coding labs, media studios and the inference platforms you buy tokens from — one stop per provider, at the model or product it would tell you to reach for first. Below them, every model in every provider’s data file, one line per provider, oldest to newest — the same data the provider reference shows, rebuilt whenever an agent updates it. Those lines are unranked: a lab that publishes no figure on an independent board gets a place, never a position. Every provider links to its full page in the provider reference, and the Cheat sheet view under the map shows the cheat sheet of whichever provider you select — every model and product.

Every station carries the figure that put it there, the board that figure came from, and the date the board was read. Where a number is the maker’s own rather than an independent board’s, it is flagged.