Grok 4.5 Ranks 10th. Cursor Is at $4 Billion ARR. That’s Not a Contradiction

by

Cursor has built a machine for not needing the best model. The Router sends hard work to frontier models and everything else to efficient ones, delivering frontier quality at roughly 60% lower cost. The swarm research pairs a frontier planner with Composer workers and hits up to 100% on complex tasks like the SQLite rebuild at a fraction of the spend. The harness produced order-of-magnitude drops in tool errors.

Read those three together and the architecture is explicit. Anthropic and OpenAI are components inside Cursor’s product, and Cursor takes the margin on volume. Being 10th doesn’t matter when you’re the layer deciding which model gets called.

And that’s exactly why Composer 3 matters more than Grok 4.6 or 4.7. In the current swarm design, the expensive planner role is filled by competitors’ models — meaning Cursor’s cost of goods is a check written to Anthropic. Composer 3, trained from scratch with ~10× compute on Colossus 2, is a bid to take the planner role in-house. If it works, Cursor stops paying its biggest suppliers and its gross margin structurally re-rates. That is the single largest financial event available to this company.

Cursor has built a business that doesn’t require winning it — while quietly spending 10× compute trying to win it anyway.

Battle for Developers by Providing More Value to Developers

On July 22, Cursor Router released with intelligent Auto-mode routing (Intelligence/Balance/Cost optimization) that delivers frontier quality at ~60% lower cost by sending hard work to frontier models and the rest to efficient ones (Grok/Composer). They doubled usage limits on individual/Teams plans for Grok/Composer/Cursor models.