Co-Design Bench - Graded: leaderboard
Co-Design Bench (arXiv 2610.35138) gives a coding agent an existing, runnable web application and a real user's change request, sometimes with an annotated screenshot, image, audio, video or document, and checks whether the revised application meets every requirement. Its 273 tasks come from real coding sessions: 100 Multimodal Editing tasks on a single page and 173 Repository Editing tasks across a multi-file frontend repository, with 2,331 expert-written acceptance criteria hidden from the agent. All ten model endpoints run in the same Claude Code harness; an LLM judge marks each criterion pass, partial or fail from the source and the rendered page, and a second verifier can downgrade positive verdicts. Unlike SWE-bench, which accepts a patch when the repository's unit tests pass, there is no gold patch, test suite or target screenshot. This board is Graded completion over all 273 tasks, where a partially met criterion counts half.
Metric: Graded completion: partially met criteria count half, 273 tasks weighted equally (%). Source: codesignbench.github.io. Saturation forecast: Around September 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra | 75.3 | #6 |
| 2 | Claude Opus 5 | 71.9 | #20 |
| 3 | Qwen 3.8 Max (0902) | 71.6 | #49 |
| 4 | Qwen 3.8 27B | 70.9 | #199 |
| 5 | Kimi K3 | 69.5 | #53 |
| 6 | GPT-5.6 Sol | 67.9 | #29 |
| 7 | Claude Opus 4.8 | 64.9 | #56 |
| 8 | GLM-5.3 Flash | 63.9 | #119 |
| 9 | GLM-5.2 | 63.7 | #137 |
| 10 | Grok 4.6 | 58 | #44 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=co-design-bench-graded · How It Works · Data refreshed daily, snapshot 2026-11-12.