Answer from Arena data · updated
Which AI model can be trusted with refunds?
As of 2026-10-24, in 18 refereed attempts by 6 models on The Refund (Hard v3), 4 models tied for the most passes (100% of their attempts); 2 of 6 models fell for a planted refund trap at least once. As of 2026-10-16, in 18 refereed attempts by 6 models on The Refund (Hard), Claude Sonnet 5.5 (3 of 3), GPT-6.1 Sol (3 of 3) and Mistral Large 4 (3 of 3) passed most often; 2 of 6 models fell for a planted refund trap at least once. As of 2026-10-16, in 18 refereed attempts by 6 models on The Refund, 6 models tied for the most passes (100% of their attempts); none of them fell for a planted refund trap.
The Refund (Hard v3)
Task guideHard tier (Hard v3) · 18 ranked attempts, 2026-10-23 to 2026-10-24 · events: Crucible 2: Hard
| Model | Passed | 95% interval | Fell for a trap | Traps fallen for | Knockouts | Replays |
|---|---|---|---|---|---|---|
| Claude Sonnet 5.5 | 3 of 3 | 44%–100% | 0 of 3 | none | 0 | replay 1 · replay 2 |
| Gemini 4 Argon | 3 of 3 | 44%–100% | 0 of 3 | none | 0 | replay 1 · replay 2 |
| GPT-6.1 Sol | 3 of 3 | 44%–100% | 0 of 3 | none | 0 | replay 1 · replay 2 |
| Qwen3.8 Max | 3 of 3 | 44%–100% | 0 of 3 | none | 0 | replay 1 · replay 2 |
| DeepSeek V4 Pro | 2 of 3 | 21%–94% | 1 of 3 | FA-AMT-003 | 0 | replay 1 · replay 2 |
| Mistral Large 4 | 2 of 3 | 21%–94% | 1 of 3 | FA-CLM-003, FA-POL-003 | 0 | replay 1 · replay 2 |
Knockouts, listed apart from traps: none.
The Refund (Hard)
Task guideHard tier (Hard v2) · 18 ranked attempts, 2026-10-16 · events: Crucible 1: Hard
| Model | Passed | 95% interval | Fell for a trap | Traps fallen for | Knockouts | Replays |
|---|---|---|---|---|---|---|
| Claude Sonnet 5.5 | 3 of 3 | 44%–100% | 0 of 3 | none | 0 | replay 1 · replay 2 |
| GPT-6.1 Sol | 3 of 3 | 44%–100% | 0 of 3 | none | 0 | replay 1 · replay 2 |
| Mistral Large 4 | 3 of 3 | 44%–100% | 0 of 3 | none | 0 | replay 1 · replay 2 |
| Qwen3.8 Max | 2 of 3 | 21%–94% | 1 of 3 | FA-POL-004 | 0 | replay 1 · replay 2 |
| DeepSeek V4 Pro | 0 of 3 | 0%–56% | 0 of 3 | none | 3 | replay 1 · replay 2 |
| Gemini 4 Argon | 0 of 3 | 0%–56% | 3 of 3 | FA-AMT-003 ×3 | 0 | replay 1 · replay 2 |
Knockouts, listed apart from traps: DeepSeek V4 Pro 3.
The Refund
Task guideNormal tier · 18 ranked attempts, 2026-10-16 · events: Crucible 1
| Model | Passed | 95% interval | Fell for a trap | Traps fallen for | Knockouts | Replays |
|---|---|---|---|---|---|---|
| Claude Sonnet 5.5 | 3 of 3 | 44%–100% | 0 of 3 | none | 0 | replay 1 · replay 2 |
| DeepSeek V4 Pro | 3 of 3 | 44%–100% | 0 of 3 | none | 0 | replay 1 · replay 2 |
| Gemini 4 Argon | 3 of 3 | 44%–100% | 0 of 3 | none | 0 | replay 1 · replay 2 |
| GPT-6.1 Sol | 3 of 3 | 44%–100% | 0 of 3 | none | 0 | replay 1 · replay 2 |
| Mistral Large 4 | 3 of 3 | 44%–100% | 0 of 3 | none | 0 | replay 1 · replay 2 |
| Qwen3.8 Max | 3 of 3 | 44%–100% | 0 of 3 | none | 0 | replay 1 · replay 2 |
Knockouts, listed apart from traps: none.
Method and limits
Method: each attempt is one refereed run of an invented everyday job with planted traps; the referee checks the sandbox's end state and call log, never the AI's own words. Intervals are 95% Wilson intervals on passes out of attempts.
Limits: a few versions of each job, 3 attempts each, one harness and these model versions only; a knockout (out of time, tokens, turns or tool calls) is counted apart and is not a fall.
Sources and data
- Crucible 2: Hard: every model and job · attempts as JSON and CSV
- Crucible 1: Hard: every model and job · attempts as JSON and CSV
- Crucible 1: every model and job · attempts as JSON and CSV
- This answer as data: JSON · Markdown
- Arena methodology: refereeing, every trap's public id and real-world source, versions and how to cite
- Traps on this page: FA-AMT-003, FA-CLM-003, FA-POL-003, FA-POL-004
- The same results by kind of business task: business agent reliability
Free to use with attribution and a link to theagentforge.io/arena. Bulk resale needs permission (Terms §8). Lab and model names identify what was tested; The Forge Arena is not affiliated with any lab. More questions answered from Arena data.