Next.js Agent Evals
2026-09-25 · Framework eval · pass@4 · 2026-09-25
- What it measures
- Models do real Next.js tasks. The current board is the agents that are still rerun. The older Previously measured rows are not folded into this column.
- How
- The official Vercel / Next.js evals. Success is pass@4: one pass in four attempts counts. The baseline has no AGENTS.md. The other arm gets the bundled Next.js docs. The columns stay separate.
- What the score means
- The percent is pass@4 on one task set. Rank follows the no-AGENTS.md baseline and keeps the source page's order. The change is with AGENTS.md minus baseline.
- Caveats
- Vercel writes the tasks and grades them. This measures a model together with Next.js, not one framework against another. The front sits above 90%, near the ceiling. Pixel Canary (OpenCode) is 5th on the source page, 90% baseline and 97% with AGENTS.md, cost N/A. This site has no vendor identity for that model, so it has no model page. Its rank stays the source rank.
- Updates
- The source page dates the last run 2026-09-25. This page synced the current board on 2026-10-10 and left out Previously measured.
- Source
- nextjs.org/evals
Baseline board
| # | Model | Baseline | Cost per run |
|---|---|---|---|
| 1 | Claude Opus 5.5 high · Claude Code | 97.0 | $0.234 |
| 2 | GPT 6 Sol high · Codex | 97.0 | $0.244 |
| 3 | Claude Fable 5.1 high · Claude Code | 97.0 | $0.722 |
| 4 | Grok 4.7 OpenCode | 94.0 | $0.109 |
| 6 | Gemini 3.8 Flash OpenCode | 90.0 | $0.605 |
| 7 | GPT-6 Astra high · Codex | 90.0 | $0.891 |
| 8 | Kimi K3 OpenCode | 84.0 | $0.281 |
| 9 | GLM 5.2 OpenCode | 81.0 | $0.152 |
| 10 | Claude Sonnet 5 Claude Code | 81.0 | $0.386 |
| 11 | Cursor Composer 2.5 Cursor | 81.0 | $0.062 |
| 12 | MiniMax M3 OpenCode | 81.0 | $0.053 |
| 13 | GPT 6 Luna high · Codex | 77.0 | $0.011 |
| 14 | Kimi K2.7 Code OpenCode | 74.0 | $0.139 |