Provisional: 1 test run. Every score below comes from a single run of our own test. See how we test.
The test
Answer New Zealand tax questions citing only Inland Revenue (IRD) pages, within a word limit. We opened every cited link.
Results
| Tool | Plan tested | Score |
|---|---|---|
| Claude | Opus, paid Max plan | 86/100 Provisional: 1 test run |
| Perplexity | Free plan, default model | 84/100 Provisional: 1 test run |
| ChatGPT | Default "Instant" model | 75/100 Provisional: 1 test run |
Claude: 86/100 (Provisional: 1 test run)
Cited only IRD pages and was the most upfront about what it could not verify. One cited link returned 404 and another failed to connect.
| Criterion | Points |
|---|---|
| Citations | 23/30 |
| Accuracy | 22/25 |
| Uncertainty | 19/20 |
| Concision | 12/15 |
| Friction | 10/10 |
Perplexity: 84/100 (Provisional: 1 test run)
All 14 cited links opened; none invented. Its source list included business.govt.nz pages despite the IRD-only brief.
| Criterion | Points |
|---|---|
| Citations | 24/30 |
| Accuracy | 21/25 |
| Uncertainty | 19/20 |
| Concision | 12/15 |
| Friction | 8/10 |
ChatGPT: 75/100 (Provisional: 1 test run)
Every cited link opened, but it ran well over the word limit and overstated that everything was "verified".
| Criterion | Points |
|---|---|
| Citations | 24/30 |
| Accuracy | 21/25 |
| Uncertainty | 12/20 |
| Concision | 8/15 |
| Friction | 10/10 |
Caveats
Accuracy was judged against the cited IRD pages; we did not check every claim independently. We plan a second run of every tool before treating any ranking as settled.