RefCraft - Reference Following: leaderboard
RefCraft (2026): reference-driven web interface implementation; B is keypoint completion, C core functionality and R reference following. Overall is the mean of each task’s minimum of B, C and R, not the minimum of their aggregate means. Nine model agents share the published protocol; provider token accounting is not a capability score. Score (%).
Metric: Score (%). Source: wangyuchi369.github.io. Saturation forecast: Around 2029. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.6 | 66.25 | #25 |
| 2 | Gemini 3.1 Pro (Preview) | 65 | #58 |
| 3 | Qwen 3.8 Max | 62.61 | #32 |
| 4 | Claude Fable 5 | 61.44 | #15 |
| 5 | Kimi K3 | 61.33 | #31 |
| 6 | GPT-6 Astra | 57.93 | #6 |
| 7 | GLM-5.3 | 55.77 | #39 |
| 8 | DeepSeek V4 Pro | 53.46 | #101 |
No result here: #4 Claude Opus 5.5, #7 Gemini 4 Argon, #8 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=refcraft-reference-following · How It Works · Data refreshed daily, snapshot 2026-10-21.