Ranking run 2026-05-11
Claude Opus 4.7 leads with 75/100 across all domains. Tax law: Claude Opus 4.7 (69) · Medicine: GPT-5 (84) · Law: Claude Opus 4.7 (79) · Business law: GPT-5 (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 | 17 / 26 | 75 | 72–78 | -1 | 55 | 1,42 | — |
| 2 | GPT-5 | 0 | 29 / 51 | 70 | 66–74 | -3 | 61 | 0,73 | — |
| 3 | Gemini 2.5 Pro | 0 | 20 / 24 | 47 | 43–51 | +3 | 68 | 4,70 | — |
| 4 | Mistral Large 2 | 0 | 19 / 88 | 43 | 38–48 | -2 | 40 | 4,30 | — |
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 · 24 Items
In Tax law, Claude Opus 4.7 (top score) contradicts a stored key fact on average 0,7 times per answer and adds avg 1,7 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 2,0 per answer. Relative to answer length: 1,86 per 1000 output tokens, lowest GPT-5 at 0,97.
| # | 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 | 69 | 63–76 | -2 | 0,7 | 1,7 | 2,4 | 100 % | 1,86 | 79 % |
| 2 | GPT-5 | 51 | 41–60 | -4 | 1,5 | 1,3 | 2,8 | 96 % | 0,97 | 70 % |
| 3 | Gemini 2.5 Pro | 31 | 24–38 | -1 | 1,4 | 0,7 | 2,0 | 96 % | 8,65 | 65 % |
| 4 | Mistral Large 2 | 11 | 8–15 | +2 | 2,4 | 1,6 | 4,0 | 100 % | 8,22 | 62 % |
Medicine · 27 Items
Ranks 1–2 statistically tied (95 % CI overlaps rank 1)
In Medicine, GPT-5 (top score) contradicts a stored key fact on average 0,3 times per answer and adds avg 0,7 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 0,7 per answer. Relative to answer length: 0,72 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 | 84 | 80–87 | +2 | 0,3 | 0,7 | 1,0 | 100 % | 0,72 | 77 % |
| 2 | Claude Opus 4.7 | 82 | 76–86 | -1 | 0,3 | 1,0 | 1,3 | 100 % | 1,23 | 79 % |
| 3 | Mistral Large 2 | 67 | 62–70 | +3 | 0,6 | 0,7 | 1,2 | 100 % | 2,05 | 66 % |
| 4 | Gemini 2.5 Pro | 53 | 48–59 | -1 | 0,4 | 0,3 | 0,7 | 67 % | 3,01 | 71 % |
Law · 24 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 0,3 times per answer and adds avg 0,7 claims the judge panel flags as false or unsupported; 79 % of answers carry at least one flag. The lowest value is Gemini 2.5 Pro at avg 0,4 per answer. Relative to answer length: 1,01 per 1000 output tokens, lowest GPT-5 at 0,53.
| # | 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 | 79 | 73–84 | +2 | 0,3 | 0,7 | 0,9 | 79 % | 1,01 | 78 % |
| 2 | GPT-5 | 71 | 66–76 | -2 | 0,3 | 0,7 | 0,9 | 75 % | 0,53 | 72 % |
| 3 | Gemini 2.5 Pro | 55 | 49–61 | +4 | 0,3 | 0,2 | 0,4 | 25 % | 1,85 | 70 % |
| 4 | Mistral Large 2 | 47 | 42–52 | -7 | 0,6 | 0,7 | 1,3 | 92 % | 4,25 | 60 % |
Business law · 24 Items
Ranks 1–2 statistically tied (95 % CI overlaps rank 1)
In Business law, GPT-5 (top score) contradicts a stored key fact on average 0,5 times per answer and adds avg 1,1 claims the judge panel flags as false or unsupported; 100 % of answers carry at least one flag. Relative to answer length: 0,60 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 | 74 | 67–80 | -5 | 0,5 | 1,1 | 1,6 | 100 % | 0,60 | 63 % |
| 2 | Claude Opus 4.7 | 69 | 65–75 | -2 | 0,7 | 1,9 | 2,5 | 100 % | 1,42 | 62 % |
| 3 | Gemini 2.5 Pro | 46 | 38–54 | +8 | 1,1 | 0,9 | 2,0 | 92 % | 5,06 | 57 % |
| 4 | Mistral Large 2 | 45 | 37–54 | -7 | 1,3 | 1,9 | 3,1 | 100 % | 3,81 | 51 % |
Raw data of this run
raw.jsonl and progress.jsonl are the public streams with response texts stripped (contamination protection, methodology §7).