Cursor Rebuilt SQLite With Agents for 1,339 Dollars

Cursor had a team of agents rebuild SQLite in Rust using only the 835-page manual. No source code, no test suites, no internet. The result had to pass a held-out sqllogictest suite, and it did, per Cursor's writeup from researcher Wilson Lin published July 20.

The engineering result is the headline, but the cost spread is the number worth sitting with. The cheapest model configuration finished the job for 1,339 dollars. Running everything on GPT-5.5 hit 10,565 dollars. Same task, same passing outcome, 7.9 times the bill.

The routing logic behind that gap is simple: reserve frontier models for the high-judgment moments, planning, architecture, deciding what "done" means, and let cheap models handle bulk execution once the plan is fixed. Translating a settled plan into 400 files of Rust doesn't need the expensive model touching every line.

That's a real update to how I'd assumed this worked. My instinct running a scrappier version of this in my own projects was that quality collapses if the expensive model isn't everywhere. Cursor's data says otherwise: once the interfaces are pinned, the judgment already happened, and you're paying frontier prices for what amounts to transcription.

A separate study adds a complication. A UC Berkeley and MIT-IBM team ran 3.1 million LLM rollouts across 30 harness configurations on 12 model-problem pairs and found no fixed harness reliably wins. Harness choice, per that paper, is a hyperparameter you tune per problem, not an architecture you settle once.

What to actually do

Instrument token spend per phase in your own work, find the plan/execute boundary, then run one task twice, once with your default model mix and once with cheap models below that boundary. Diff the results. Cursor flags its own caveat: SQLite-from-manual has a spec and an oracle most real tasks lack. Whether the savings hold when the spec is a Slack thread isn't something either study answers.