Cheap coding models did the work while a smart model planned it

Cursor's agent swarm rebuilt SQLite in Rust from docs alone, with expensive planners directing cheap workers at a fraction of solo-model cost.

Cursor tested an upgraded AI agent swarm by asking it to rebuild SQLite in Rust using only an 835-page manual — no source code, test suites, binary, or internet access. Every configuration of the new system reached 100 percent on the sqllogictest benchmark, which contains millions of SQL queries with known answers. The design splits agents into planners running frontier models and workers running cheaper, faster models.

The cost spread is the headline. Total run costs ranged from $1,339 for an Opus planner paired with the cheaper Composer 2.5 worker, up to $10,565 for GPT-5.5 running solo. In the GPT-5.5 run, worker tasks alone cost $9,373; in the Opus hybrid, the entire worker fleet cost $411. Composer 2.5 is priced at $0.50 per million input tokens and $2.50 per million output. The swarm hit roughly 1,000 commits per second, forcing Cursor to build its own version control to manage it.

If frontier models are needed only to plan and cheap models can execute, the marginal cost of producing working software drops sharply. That pushes software creation toward something closer to a commodity, widening who can build serious tools.

The caveats are real: these are Cursor's own tests and reporting. Planner quality still mattered — the Fable 5 setup cost more overall despite using fewer planning tokens — so the cheap-worker approach is not a free lunch. Independent replication would strengthen the claim.

Source: The Decoder