Next.js Framework Work
How good are agents at Next.js?
Run by mupt-ai · Released 2026-11-10 · 2 repositories · 56 tasks · 12 settings
Model Tradeoffs
Which coding agent works best on Next.js Framework Work? Most accurate: switchyard, 78.6% at $0.51 per task. Best for less: Dari Router, 75% at $0.35. 12 model settings scored on 56 tasks from the merged pull requests of 2 repositories.
All Settings
| Model | Harness | Reasoning | Access | Accuracy | Cost / Task | Frontier |
|---|---|---|---|---|---|---|
| switchyard | Codex | Default | Custom Endpoint | 78.6% | $0.51 | Yes |
| Claude Opus 5.5 | Claude Code | Max | API Key | 78.6% | $7.78 | No |
| Dari Router | Pi | Default | API Key | 75% | $0.35 | Yes |
| Claude Opus 5.5 | Claude Code | Low | API Key | 75% | $0.70 | No |
| Claude Opus 5.5 | Claude Code | High | API Key | 75% | $1.27 | No |
| GPT-6.1 Sol | Codex | High | API Key | 71.4% | $0.51 | No |
| Jev Router | Pi | High | OpenRouter | 71.4% | $3.69 | No |
| DeepSeek V4.1 Flash 1017 | Codex | High | Vercel AI Gateway | 67.9% | $0.18 | Yes |
| DeepSeek V4.1 Flash 1017 | Pi | High | Vercel AI Gateway | 64.3% | $0.16 | Yes |
| Claude Haiku 5.5 | Claude Code | High | API Key | 60.7% | $0.26 | No |
| Claude Haiku 5.5 | Claude Code | Low | API Key | 57.1% | $0.059 | Yes |
| GPT-6 Luna | Codex | High | API Key | 55.4% | $0.034 | Yes |
Repositories
- dubinc/dub.link 18 tasks. Most accurate here: Claude Opus 5.5, 88.9%.
- calcom/cal.diy 38 tasks. Most accurate here: switchyard, 78.9%.
Model Summaries
For routine work, Claude Opus 5.5 · low reasoning · Claude Code, Claude Opus 5.5 · high reasoning · Claude Code, and GPT-6.1 Sol · high reasoning · Codex are the strongest choices. For large, difficult features, Claude Opus 5.5 · max reasoning · Claude Code is the strongest option on its mostly separate task set, while DeepSeek V4.1 Flash 1017 · high reasoning · Pi is capable but less consistent; Claude Haiku 5.5 · high reasoning · Claude Code, DeepSeek V4.1 Flash 1017 · high reasoning · Codex, and GPT-6 Luna · high reasoning · Codex need closer review. Higher reasoning does not consistently improve results on the same work, and the strongest max-reasoning results come from mostly different tasks.
Claude Opus 5.5 · Claude Code · Max
It is a strong choice for difficult, large multi-file features spanning UI, storage and external integrations. It usually delivers end-to-end behavior across those layers, although routine and small changes were scarcely exercised. Review data and access semantics and check existing behavior on ambitious changes, since failures can leave requirements unmet or disrupt existing behavior. Time limits are a recurring risk, including on some attempts that otherwise passed.
Claude Opus 5.5 · Claude Code · Low
It is reliable on routine bug fixes and medium-sized server-side refactors, including changes spread across many files. It also handles many larger changes, but difficult UI refactors and end-to-end features deserve closer review. Check that requested behavior is complete throughout the interface and supporting layers, since failures tend to leave parts of the change wrong rather than run out of time.
Claude Opus 5.5 · Claude Code · High
It handles routine fixes and medium-sized refactors reliably, including work spread across files. Larger, difficult framework and UI changes are less consistent, although it can complete substantial multi-file features. Check behavior across server and client boundaries and confirm that all affected UI paths were updated. Failures generally leave requested behavior wrong rather than run out of time.
GPT-6.1 Sol · Codex · High
It handles routine bug fixes and medium-sized multi-file refactors well, with some successful larger migrations. Difficult server-side refactors and features spanning UI and persisted data are less dependable. Check access-related decisions, component behavior and end-to-end completeness, since failures generally leave requested behavior incorrect rather than run out of time.
DeepSeek V4.1 Flash 1017 · Codex · High
It often handles routine fixes and medium-sized refactors and can complete difficult changes spread across many files. Feature work is possible, but hard refactors and security-sensitive changes are less predictable. Check behavior across affected paths, access boundaries and imports, since failures commonly leave requirements unmet and occasionally run out of time.
DeepSeek V4.1 Flash 1017 · Pi · High
It can deliver many hard, large multi-file features spanning UI, data persistence and external integrations. Its coverage of routine or small work is too limited to assess, and large cross-layer features are not consistently correct. Review access controls, persistence and end-to-end behavior, particularly where APIs and UI must stay in sync. Failures generally leave requested behavior missing or wrong rather than run out of time.
Claude Haiku 5.5 · Claude Code · High
It handles many routine bug fixes and server-side refactors, including some large changes spread across files, but is inconsistent even on easy work. Medium and difficult changes need close review, especially when they involve access control or several execution paths. Failures commonly leave requested behavior missing or wrong, and an occasional larger change runs out of time.
GPT-6 Luna · Codex · High
It can handle routine bug fixes and some substantial multi-file server refactors, but is uneven even on easy or medium work. It can complete large changes, though difficult features spanning interface and data need close review. Check access decisions, server-side behavior and end-to-end completeness, since failures commonly leave required behavior missing or wrong. Incomplete behavior is a greater concern than time limits in this sample.
Written by openai/gpt-6.1-terra from these results on 2026-11-10.
Results by Task
Whether each setting passed each task, with how its attempt was graded and its transcript.
Tasks
All 56 tasks, each with its instruction, tests, and solution to browse and download.