Next evaluationfor solversof agents
NextEval
NextEval is an open research project on derivative-free optimization with language-model agents under a fixed budget of function evaluations: a benchmark that scores coding agents against COBYQA, and a solver in which an agent chooses the numerical methods.
NextEval Bench
A benchmark that scores coding agents on derivative-free optimization problems against the COBYQA solver.
Open the leaderboard /solverRelease: coming soonNextEval Solver 1.0
A derivative-free optimizer in which a language-model agent chooses numerical-method campaigns, with a textbook autopilot as the fixed-policy fallback.
Read about the solverBench at a glance
Full leaderboardLast updated: 2026-11-06 · snapshot web-ffdd187d
- 45agents, each a harness driving a model
- 92problems × 4 features × 3 repeats
- 1,104cases per agent, each compared with COBYQA
First six of 45 agents by pair score against COBYQA (4-task mean). COBYQA = 0.500; above means ahead of COBYQA. The beat-COBYQA rate is the share of cases in which the agent reaches every tolerance no later than COBYQA and at least one strictly earlier.
| Rank | Agent | Harness | From COBYQA (0.500) | Pair score | Beat rate |
|---|---|---|---|---|---|
| 1 | GPT-6.1 Sol (high) | Codex | 0.726 | 38.0% | |
| 2 | GPT-6 Astra (high) | Codex | 0.706 | 37.2% | |
| 3 | Opus 5.5 (high) | Claude Code | 0.691 | 37.8% | |
| 4 | GPT-6.1 Sol (medium) | Codex | 0.690 | 33.8% | |
| 5 | Fable 5.1 (high) | Claude Code | 0.680 | 34.5% | |
| 6 | Hy4 (high) | Claude Code | 0.674 | 34.1% |
Views: Leaderboard · Matrix · Waffle · Profiles · Pareto · Methodology
Code and results
- Code
- github.com/NextEval, the NextEval organization on GitHub.
- Results
- The published results, with the evaluation history of every problem and repeat, are in NextEval Bench; how they are scored is in its methodology.
- Website
- NextEval/nexteval.github.io, the source of these pages.