Researchers have built an AI system that evolves its own code more efficiently - and a lot more cheaply.
A new paper describes FrugalEvo, a framework for LLM-guided evolutionary programming, the technique that made headlines with AlphaEvolve. Instead of running every step on one expensive model, FrugalEvo splits the labor: a stronger, costlier LLM proposes solution strategies, while a cheaper model writes and refines the actual code. The team also built a cache-efficient harness that reuses shared prompt prefixes across evolution steps to cut redundant computation. Tested on 10 math and systems optimization tasks plus 10 algorithmic tasks from ALE-Bench-Lite, FrugalEvo matched or beat existing baselines including OpenEvolve, ShinkaEvolve, AdaEvolve, and EvoX.
The headline number is cost. On the classic circle-packing problem, FrugalEvo hit state-of-the-art results for $1.68 using GPT-5.6 Terra and Luna, and for $0.55 using GLM-5.3 and its Flash variant. Multi-agent approaches like CORAL and SwarmResearch needed roughly $50 to reach comparable results - call it a 30x to 90x cost reduction for the same output.
Evolutionary code search has mostly been treated as a brute-force spending contest since AlphaEvolve made headlines; this paper is a reminder that the real constraint was never compute, it was budget discipline.