WARC-Bench (192-task test split)
Agents
Uniphore's WARC-Bench (ICLR 2026) tests multimodal computer-use agents on GUI subtasks, short-horizon interactions with real UI widgets such as choosing a date in a date picker, navigating a dropdown or scrolling a container to read a value, on real websites replayed from Web Archive (WARC) snapshots in a sandboxed Chromium browser. Every task has a programmatic grader of the end state. This entry is the 192-task whitelisted test split (of the 438-task release) run with Uniphore's screenshot-only Subtask Vision Agent at 20 steps per task; the score is the share of tasks completed, averaged over three complete runs.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 90.3 |
| GPT-6 Astra | 89.1 |
| GPT-6.1 Sol | 86.3 |
| Kimi K3 | 85.9 |
| Gemini 4 Argon | 76.6 |
| DeepSeek V4.1 Flash 1017 | 70.5 |
| Muse Spark Open | 65.5 |
| Grok 4.7 | 35.6 |