BENCHMARK RESULTS
Leaderboard
How well can coding agents build search engines?
Compare results. Follow the code. Read the evidence.
Evaluation snapshot2026-10-31Fixed snapshot ↗
| Rank | Model | Agent | Tasks | ||||
|---|---|---|---|---|---|---|---|
| 1 | GPT-6 Luna | Codex | 16.4 | 47 / 47 | 1.07B | $20.51 | |
| 2 | DeepSeek V4.1 Flash 1017 | Codex | 15.5 | 47 / 47 | 2.42B | $20.92 | |
| 3 | MiMo-V2.6-Flash | Codex | 4.7 | 47 / 47 | 528.4M | $9.76 |
No matching models. Try a different search.
Score vs. cost
Pareto frontierNo matching models with complete cost data.
All 47 tasks · one selected evaluation per model and taskDownload results JSON ↓
Score: equal-weight task average (0–100). Tokens: total coding input (including cache) + output.
Cost is estimated from recorded agent tokens at official API rates in USD (2026-11-01; DeepSeek off-peak pricing).