Provisional: 1 test run. Every score below comes from a single run of our own test. See how we test.
The test
Sort a small-business bank CSV, find three planted errors (a duplicate, a sign-flipped refund, an owner drawing) and give a corrected net.
Results
| Tool | Plan tested | Score |
|---|---|---|
| ChatGPT | "Instant", signed-in account, plan not visible | 93/100 Provisional: 1 test run |
| Claude | Opus 5.5 (Medium), paid Max plan | 90/100 Provisional: 1 test run |
ChatGPT: 93/100 (Provisional: 1 test run)
Caught all 3 traps and reached the expected corrected net of 3,436.70. Avoided tax commentary, but showed the corrected figure only in a closing sentence.
| Criterion | Points |
|---|---|
| Accuracy | 35/35 |
| Error detection | 25/25 |
| Clarity | 13/15 |
| Safety | 15/15 |
| Friction | 5/10 |
Claude: 90/100 (Provisional: 1 test run)
Caught all 3 traps and reached 3,436.70 with a clear as-recorded vs corrected table. Lost points for unrequested tax commentary.
| Criterion | Points |
|---|---|
| Accuracy | 34/35 |
| Error detection | 25/25 |
| Clarity | 14/15 |
| Safety | 12/15 |
| Friction | 5/10 |
Caveats
The gap is within likely run-to-run variation, so this is not a firm ranking. Neither run was a confirmed free-tier run. We plan a second run of every tool before treating any ranking as settled.