Cursor's agent swarm suggests cheaper models can handle most coding when frontier models plan the work
Quick Answer
Cursor's agent swarm demonstrates that cheaper models can effectively handle coding tasks when paired with powerful frontier models for planning, achieving 1,000 commits per second and reducing codebase size by up to 85%.
Quick Take
The hybrid approach significantly cut costs, with worker models costing as little as $411 compared to $9,373 for GPT-5.5, while maintaining high performance in benchmarks.
Key Points
- Cursor's new swarm achieved 1,000 commits per second, a significant improvement over previous models.
- The hybrid system reduced codebase size by up to 85% while maintaining high performance.
- Worker models accounted for at least 69% of tokens, drastically lowering operational costs.
- The cheapest configuration used Opus and Composer, costing $411 compared to $9,373 for GPT-5.5.
- Multiple review perspectives improved reliability, with no single view catching all issues.
DeepSignal Analysis
What happened
Cursor's agent swarm utilizes a hybrid approach where planner agents use advanced models to break tasks down, while cheaper worker models execute these tasks. This system achieved 1,000 commits per second and reduced codebase size by up to 85%. The cost of using cheaper worker models was significantly lower, with worker costs at $411 compared to $9,373 for GPT-5.5.
Key evidence
- Cursor's new swarm reached 1,000 commits per second, necessitating a custom version control system due to unique failure modes.
- The new system outperformed the old one in every configuration, achieving scores between 73 and 85 percent, while the old system scored between 11 and 77 percent.
- In the Opus hybrid configuration, the total cost was $1,339, while the GPT-5.5 run alone cost $10,565, highlighting a significant cost difference.
Why it matters
The findings suggest that a combination of powerful planning models and cheaper execution models can optimize coding tasks, leading to substantial cost savings and efficiency improvements. This hybrid approach may redefine how software development is approached, particularly in environments where rapid iteration and cost management are critical. The ability to maintain high performance while significantly reducing codebase size could also influence future software engineering practices.
Source Excerpt
Cursor asked its upgraded agent swarm and its predecessor to rebuild SQLite in Rust using only the documentation, with no source code or internet access. Every configuration of the new system, which separates planners from workers, eventually scored 100 percent on the test suite. The old swarm choked on merge conflicts of its own making.
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from The Decoder
See more →
An AI model programmed nonstop for 19 days on a single MirrorCode task that cost $2,600 to run
Epoch AI's MirrorCode benchmark reveals Claude Opus 4.7 as the leader with a 56% solve rate, reconstructing a 16,000-line toolkit in 14 hours. Despite this, all models tested struggle with the most complex tasks, highlighting limitations in current AI capabilities. The single task consumed $2,600 over 19 days, raising questions about cost-effectiveness in AI development.
