Agent and model comparison with imge session replay
Berkay Demirbas··2 min readOn a local Studio Store demo, the task was to click Add to cart without choosing a size, select Medium, click Add to cart again, and report the first failure and the final cart count.
The page has two buttons with the same accessible name. The first is a trap. Each run used the imge MCP browser tools and started from a fresh browser session.
| Client | Model | Wall time | Frames | Cart |
|---|---|---|---|---|
| Codex | gpt-6-luna, low effort | ~20 s | 5 | 1 |
| Cursor | Composer 2.5 | ~58 s | 9 | 0 |
| Cursor | GPT-5.3 Codex Low | 31.9 s | 4 | 0 |
| Cursor | Claude Sonnet 5 Thinking High | 51.6 s | 5 | 1 |
| Cursor | Grok 4.7 Low | 41.2 s | 4 | 0 |
| Cursor | Auto | 104.2 s | 4 | 0 |
| Cursor | Gemini 3 Flash | 32.6 s | 5 | 1 |
| Cursor | GLM 5.2 High | 39.7 s | 6 | 0 |
| Cursor | DeepSeek V4.1 Flash Low | 19.7 s | 5 | 1 |
The first click failed with “Choose a size first” in every completed run. The cart ended at 1 when the run used the second button, and at 0 when it used the first. Wall time includes CLI startup, MCP initialization, model time, and browser actions. The Codex and Composer times are approximate, taken from log timestamps. The other Cursor times came from a timed runner.

Berkay Demirbas
Creator of imge