WARC-Bench: leaderboard
WARC-Bench (Uniphore, arXiv 2510.09872; test-split runs of 19 October to 10 November 2026) tests whether an agent can operate single web interface components: 192 short GUI subtasks such as picking a date in a calendar widget, choosing from a dropdown, scrolling a container to read a value or completing a multi-step form, replayed from web archive (WARC) snapshots of real sites. Unlike WebArena or Online-Mind2Web, which grade whole multi-page tasks, it isolates the short interactions those tasks are built from and replays recorded pages, so every run sees the same site. The agent sees screenshots only and gets at most 20 steps and 1,800 seconds per task; a programmatic check grades the end state, 53 tasks by exact string match, with no judge model. Each model runs the split 3 times at the highest reasoning effort its API accepts, and the score is the mean of the 3 runs. Four of the models are served by Fireworks.
Metric: Tasks completed on the 192-task test split, mean of 3 runs (%). Source: github.com. Saturation forecast: Around February 2027. 9 models tracked. Updated 14 Nov 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 | Max | 90.28 | #5 |
| 2 | GPT-6 Astra | xHigh | 89.06 | #6 |
| 3 | GPT-6.1 Sol | xHigh | 86.28 | #10 |
| 4 | Kimi K3 | Max | 85.94 | #56 |
| 5 | Gemini 4 Argon | High | 76.56 | #12 |
| 6 | DeepSeek V4.1 Flash (1017) | Max | 70.49 | #41 |
| 7 | Muse Spark Open | Max | 65.45 | #81 |
| 8 | MiniMax-M3.1-Flash | Max | 43.92 | #109 |
| 9 | Grok 4.7 | xHigh | 35.59 | #47 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon.
Interactive version: theaggregate.ai/benchmark?slug=warc-bench · How It Works · Data refreshed daily, snapshot 2026-11-16.