Ranking run 2026-05-10
Claude Opus 4.7 leads with 78/100 across all domains. Tax law: GPT-5 (90) · Medicine: Claude Opus 4.7 (86) · Law: Claude Opus 4.7 (80) · Business law: Claude Opus 4.7 (74).
Overall across domains
| # | Model | top100 | 95 % CI | Δ prev. | failures | sec median/p90 | hallu band | per 1000 tok. | Δ own family |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 4.7 | 0 | 19 / 30 | 78 | 74–82 | +1 | 54 | 1,47 | — |
| 2 | GPT-5 n=16/33 | 17 | 29 / 37 | 76 | 69–83 | -2 | 81 | 0,48 | — |
| 3 | Mistral Large 2 | 0 | 18 / 370 | 45 | 38–54 | -1 | 38 | 4,30 | — |
| 4 | Gemini 2.5 Pro n=5/33 | 28 | 33 / 35 | 15 | 10–23 | -17 | 96 | 1,61 | — |
Δ prev. = difference to the previous published run; “—” = model was not in it. Hallu band: 100 = none, 0 = ≥ 4 false claims per answer. Per 1000 tok. = flagged claims per 1000 output tokens (length-normalised). Δ own family = score with all judges minus score with foreign judges only, in points; positive = own family scores more generously; “—” = not recorded before code_version 0.2.8 or no foreign judge (methodology §4, §6.5).
Tax law · 1 Items
In Tax law, GPT-5 (top score) contradicts a stored key fact on average 0,0 times per answer and adds avg 0,0 claims the judge panel flags as false or unsupported; 0 % of answers carry at least one flag. Relative to answer length: 0,00 per 1000 output tokens.
| # | Model | top100 | 95 % CI | Δ prev. | contradicts key fact | extra false/unsupp. | avg total | answers affected | per 1000 tok. | IRR |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | GPT-5 | 90 | 90–90 | — | 0,0 | 0,0 | 0,0 | 0 % | 0,00 | 71 % |
| 2 | Claude Opus 4.7 | 71 | 64–80 | +5 | 0,5 | 2,0 | 2,5 | 100 % | 1,89 | 87 % |
| 3 | Gemini 2.5 Pro | 12 | 13–13 | +12 | 0,0 | 0,3 | 0,3 | 100 % | 3,92 | 100 % |
| 4 | Mistral Large 2 | 12 | 8–15 | +3 | 2,5 | 1,3 | 3,8 | 100 % | 8,36 | 57 % |
Medicine · 9 Items
Ranks 1–2 statistically tied (95 % CI overlaps rank 1)
In Medicine, Claude Opus 4.7 (top score) contradicts a stored key fact on average 0,3 times per answer and adds avg 0,8 claims the judge panel flags as false or unsupported; 100 % of answers carry at least one flag. The lowest value is Gemini 2.5 Pro at avg 0,1 per answer. Relative to answer length: 1,10 per 1000 output tokens, lowest GPT-5 at 0,70.
| # | Model | top100 | 95 % CI | Δ prev. | contradicts key fact | extra false/unsupp. | avg total | answers affected | per 1000 tok. | IRR |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 4.7 | 86 | 81–92 | 0 | 0,3 | 0,8 | 1,1 | 100 % | 1,10 | 79 % |
| 2 | GPT-5 | 82 | 76–88 | -2 | 0,1 | 0,9 | 1,0 | 78 % | 0,70 | 68 % |
| 3 | Mistral Large 2 | 66 | 55–76 | -2 | 0,7 | 1,0 | 1,7 | 100 % | 2,70 | 67 % |
| 4 | Gemini 2.5 Pro | 19 | 13–29 | -24 | 0,0 | 0,1 | 0,1 | 33 % | 1,05 | 92 % |
Law · 8 Items
Ranks 1–2 statistically tied (95 % CI overlaps rank 1)
In Law, Claude Opus 4.7 (top score) contradicts a stored key fact on average 0,4 times per answer and adds avg 0,7 claims the judge panel flags as false or unsupported; 88 % of answers carry at least one flag. The lowest value is Gemini 2.5 Pro at avg 0,0 per answer. Relative to answer length: 1,23 per 1000 output tokens, lowest Gemini 2.5 Pro at 0,00.
| # | Model | top100 | 95 % CI | Δ prev. | contradicts key fact | extra false/unsupp. | avg total | answers affected | per 1000 tok. | IRR |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 4.7 | 80 | 72–88 | +2 | 0,4 | 0,7 | 1,1 | 88 % | 1,23 | 80 % |
| 2 | GPT-5 | 65 | 55–77 | -5 | 0,0 | 0,5 | 0,5 | 67 % | 0,29 | 58 % |
| 3 | Mistral Large 2 | 49 | 41–57 | -4 | 0,8 | 0,4 | 1,2 | 75 % | 3,52 | 56 % |
| 4 | Gemini 2.5 Pro | 7 | 7–7 | -31 | 0,0 | 0,0 | 0,0 | 0 % | 0,00 | 89 % |
Business law · 8 Items
In Business law, Claude Opus 4.7 (top score) contradicts a stored key fact on average 0,5 times per answer and adds avg 2,2 claims the judge panel flags as false or unsupported; 100 % of answers carry at least one flag. Relative to answer length: 1,52 per 1000 output tokens.
| # | Model | top100 | 95 % CI | Δ prev. | contradicts key fact | extra false/unsupp. | avg total | answers affected | per 1000 tok. | IRR |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 4.7 | 74 | 65–83 | -4 | 0,5 | 2,2 | 2,7 | 100 % | 1,52 | 74 % |
| 2 | Mistral Large 2 | 53 | 41–62 | +3 | 1,4 | 1,8 | 3,2 | 100 % | 3,74 | 59 % |
Raw data of this run
raw.jsonl and progress.jsonl are the public streams with response texts stripped (contamination protection, methodology §7).