Next.js Agent Evals
How well AI coding agents perform real Next.js tasks.
Current evals
Latest agents, rerun whenever the evals change.
Agent | |||||
|---|---|---|---|---|---|
GPT 6.1 Terra (high) | Codex | 169.36s | $0.080 | 100% | 100% |
Claude Opus 5.5 (high) | Claude Code | 275.52s | $0.215 | 100% | 100% |
GPT 6 Sol (high) | Codex | 284.24s | $0.247 | 100% | 100% |
Gemini 4 Argon | OpenCode | 269.59s | $0.445 | 100% | 100% |
Claude Fable 5.1 (high) | Claude Code | 244.19s | $0.676 | 100% | 100% |
Grok 4.7 | OpenCode | 340.47s | $0.174 | 96% | 96% |
Gemini 3.8 Flash | OpenCode | 441.39s | $0.615 | 93% | 100% |
GPT 6 Astra (high) | Codex | 230.49s | $0.842 | 93% | 100% |
Pixel Canary | OpenCode | 1005.28s | N/A | 93% | 100% |
GPT 6 Luna (high) | Codex | 277.23s | $0.011 | 86% | 96% |
Kimi K3 | OpenCode | 350.63s | $0.270 | 86% | 100% |
MiniMax M3 | OpenCode | 234.46s | $0.050 | 82% | 93% |
Cursor Composer 2.5 | Cursor | 266.93s | $0.060 | 82% | 96% |
GLM 5.2 | OpenCode | 378.28s | $0.147 | 82% | 100% |
Claude Sonnet 5 | Claude Code | 258.32s | $0.349 | 82% | 100% |
Kimi K2.7 Code | OpenCode | 318.37s | $0.134 | 75% | 93% |
Previously measured
Superseded agents. Last results, dated — not rerun.
Agent | |||||
|---|---|---|---|---|---|
Claude Opus 5 Sep 9, 2026 | Claude Code | 360.60s | $1.90 | 96% | 100% |
Muse Spark Open Oct 16, 2026 | OpenCode | 321.96s | $0.220 | 93% | 100% |
GPT 5.6 Sol (ultra) Sep 8, 2026 | Codex | 316.09s | $0.660 | 79% | 79% |
Claude Fable 5 (high) Sep 8, 2026 | Claude Code | 256.22s | $1.70 | 79% | 79% |
Grok 4.6 Aug 13, 2026 | OpenCode | 228.25s | $0.142 | 71% | 71% |
GPT 5.4 (xhigh) Jul 17, 2026 | Codex | 212.43s | $0.233 | 68% | 71% |
GPT 5.3 Codex (xhigh) Aug 27, 2026 | Codex | 275.96s | $0.243 | 68% | 75% |
Claude Opus 4.8 May 28, 2026 | Claude Code | 158.88s | $0.293 | 68% | 75% |
Grok 4.5 Jul 8, 2026 | OpenCode | 138.20s | $0.069 | 64% | 75% |
GPT 5.5 Pro Apr 26, 2026 | Codex | 753.33s | $17.67 | 64% | 68% |
Cursor Composer 2.0 Apr 6, 2026 | Cursor | 111.22s | $0.028 | 57% | 71% |
GLM 5.1 Apr 16, 2026 | OpenCode | 244.18s | $0.071 | 57% | 75% |
Gemini 3.1 Pro Preview Apr 6, 2026 | Gemini CLI | 239.15s | $0.106 | 57% | 71% |
Claude Opus 4.6 Apr 6, 2026 | Claude Code | 181.16s | $0.191 | 57% | 75% |
Claude Opus 4.7 (max) Apr 16, 2026 | Claude Code | 136.45s | $0.389 | 57% | 75% |
Kimi K2.6 Jun 17, 2026 | OpenCode | 189.25s | $0.051 | 50% | 75% |
Gemini 3.0 Pro Preview Apr 6, 2026 | Gemini CLI | 256.44s | $0.162 | 50% | 64% |
GPT 5.2 Codex (xhigh) Apr 6, 2026 | Codex | 137.86s | $0.124 | 46% | 64% |
Claude Sonnet 4.6 Apr 6, 2026 | Claude Code | 147.04s | $0.123 | 43% | 75% |
MiniMax M2.7 Apr 6, 2026 | OpenCode | 288.67s | $0.025 | 36% | 46% |
Claude Sonnet 4.5 Apr 6, 2026 | Claude Code | 140.90s | $0.115 | 36% | 64% |
Kimi K2.5 Apr 6, 2026 | OpenCode | 133.78s | $0.018 | 18% | 46% |
* AGENTS.md points agents to bundled Next.js documentation. Both conditions can access these docs. With AGENTS.md shows the success rate with that initial guidance.
N/A marks evals a model has never run (newly added evals). They count against the success rate so all models are scored over the same eval set. Reworked evals keep a model's last result until it is rerun.
Success Rate is pass@4: an eval passes if any of four attempts passes. Infrastructure failures are discarded and rerun, never counted.
Metrics average attempts within each task and setup. Model rows give those averages equal weight; expanded rows combine available Base and AGENTS.md averages for one task. Duration is recorded run time. Cost uses public model prices at export, including cache pricing. Some older costs use text-based estimates. Missing values, Sandbox and serving costs are excluded.