Cursor's AI agent swarm rebuilt SQLite in Rust, passing all tests with an 835-page manual. A mixed setup using Opus 4.8 as planner and Composer 2.5 as workers cost $1,339 versus $20,000 for a single Frontier model, delivering the same quality at 7.9× lower cost.
Cursor had an agent swarm rebuild SQLite in Rust from scratch. An 835-page manual graded by SQL logic tests. And look at what happened when they changed the model mix. You've got Fable 5 here at the highest cost, around $20,000. Then you've got Opus 4.8 with Composer 2.5 splitting up the work. That came in at $1,339, almost $1,337. $2 off, right? The architecture is the interesting part. Frontier models act as planners. They break down the goals. Then, fast, cheaper worker agents do the implementation. The planner doesn't implement, so its context doesn't fill up with all that low-level detail. And the worker doesn't plan. It gets to spend its context on one narrow piece of work. GPT-5.5 alone burned 14.7 billion tokens. Fable 5 used 12.2 billion. But the Frontier planner and cheaper worker combination delivered near-identical quality at 7.9 times lower cost than Frontier only. Once the planner turns that ambiguity into a detailed instruction, the less expensive model can just follow it. And look at the quality bar. The new swarm hit 100% on the tests with the Fable 5 mix. 9,908 lines of code, Compared with 64,000 for the old Swarm. So you're getting about 6.5 times less code. That's the thesis here. Use the frontier model for planning and orchestration, then use lower-cost models for implementation. The way I think about this is, where does the more capable model actually earn its cost? If you've got open-ended planning, especially creative planning, there's room for it to produce a better result. Then it can orchestrate other agents to go and do the work. But if the task is already bounded, you've got a fixed spec and you know what you want, a model like Opus 4.8 can be a pretty good fit. There's less room for the extra creativity to make a difference. You want to spend more where you actually expect a better result. I'd still be curious to see Composer 2.5 on its own. How well does it handle the planning and orchestration? These are the kinds of experiments we're all running as builders. Cursor just ran one at scale and put numbers to it. That's really useful for the rest of us.
