Skip to content
NewsResearch

Cursor's agent swarm rebuilt SQLite in Rust from documentation alone, with costs ranging from $1,339 to $10,565 across four model configurations

· by Pondero Newsdesk

The short version

Cursor published a July 2026 benchmark showing its new agent swarm rebuilt the 835-page SQLite specification in Rust in four hours, with all four model configurations eventually passing 100% of the test suite and run costs varying nearly 8x between hybrid and all-frontier model pairings.

Cursor's agent swarm rebuilt SQLite in Rust from documentation alone, with costs ranging from $1,339 to $10,565 across four model configurations

Give a swarm 835 pages of database spec, no source code, no internet, and a four-hour clock. That was Cursor's benchmark. On July 20, Wilson Lin published a detailed write-up of what the team's new swarm architecture produced, and what it cost.

What happened

Cursor ran the new swarm against a task where its older architecture had collapsed: rebuilding SQLite in Rust from documentation alone. Graded against sqllogictest (the SQLite project's own suite of millions of queries), the new Grok 4.5 run hit 80% within four hours, per the post. On the same task, the older swarm accumulated more than 70,000 merge conflicts before the team paused it before its second hour.

Four model configurations ran against the new harness:

  • GPT-5.5 as both planner and worker
  • Grok 4.5 as both planner and worker
  • Opus 4.8 as planner, Composer 2.5 as worker
  • Fable 5 as planner, Composer 2.5 as worker

At the four-hour mark, new runs sat between 73% and 85%. Old runs ranged from 11% to 77% at the same point. Every new configuration eventually passed 100% of the suite.

Costs split sharply. Opus 4.8 as planner with Composer 2.5 handling execution cost $1,339 for the full run. Running GPT-5.5 for both roles cost $10,565. To see where the gap came from: the GPT-5.5 worker fleet alone was $9,373. In the Opus 4.8 hybrid, workers cost $411 total.

Why it matters

Few moments in a large coding task require frontier-level intelligence. Decomposing the goal, making major design decisions, resolving architectural trade-offs: those benefit from a powerful planner. Once the frontier model has turned ambiguity into explicit, detailed instructions, cheaper models follow them reliably. That structural insight is what Cursor's data now puts numbers on.

Operators building or buying agentic coding platforms face a concrete decision point. Model mix matters as much as model selection, and the tested hybrid configurations delivered similar test-suite results at a fraction of the all-frontier price. For teams using Cursor at Business or Enterprise scale, where swarm runs can span hours, the cost differential compounds quickly.

One notable supporting finding: Cursor built a custom version-control system to replace Git for swarm work. It handles approximately 1,000 commits per second. Earlier browser-building experiments peaked at roughly 1,000 commits per hour under Git. Under the old harness, the single hottest file in the Grok 4.5 run drew 7,771 conflicts from 1,173 distinct agents. Under the new harness, the most-contested file in the entire run saw 47.

What to watch next

Cursor flagged two open directions: a full N-by-N matrix of planner-worker model pairings (the July run compared harness versions, not every combination), and research into training models to write better context for their successor agents. For outside observers, independent replication of the planner-worker cost structure on tasks other than SQLite would confirm how broadly the economics generalize. The codebase from the solo Opus 4.8 run is public at github.com/cursor/minisqlite.

Sources