Ranking run 2026-05-16
Claude Opus 4.7 leads with 78/100 across all domains. Tax law: Claude Opus 4.7 (69) · Medicine: GPT-5 (85) · Law: Claude Opus 4.7 (85) · Business law: GPT-5 (79).
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 | 21 / 36 | 78 | 73–83 | +3 | 53 | 1,50 | — |
| 2 | GPT-5 | 0 | 30 / 44 | 73 | 68–79 | +3 | 59 | 0,74 | — |
| 3 | Gemini 2.5 Pro | 0 | 20 / 24 | 46 | 40–52 | +1 | 72 | 4,17 | — |
| 4 | Mistral Large 2 | 0 | 16 / 368 | 44 | 36–52 | -2 | 44 | 4,06 | — |
| 5 | Grok 4.3 | 0 | 9 / 17 | 41 | 33–49 | — | 69 | 4,83 | — |
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 · 8 Items
Ranks 1–2 statistically tied (95 % CI overlaps rank 1)
In Tax law, Claude Opus 4.7 (top score) contradicts a stored key fact on average 0,9 times per answer and adds avg 1,9 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: 2,24 per 1000 output tokens, lowest GPT-5 at 0,81.
| # | 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 | 61–76 | +1 | 0,9 | 1,9 | 2,8 | 100 % | 2,24 | 78 % |
| 2 | GPT-5 | 61 | 48–73 | +8 | 0,9 | 1,5 | 2,4 | 100 % | 0,81 | 80 % |
| 3 | Gemini 2.5 Pro | 26 | 16–40 | -4 | 1,3 | 0,8 | 2,0 | 88 % | 9,17 | 65 % |
| 4 | Grok 4.3 | 18 | 8–28 | — | 1,6 | 0,5 | 2,1 | 88 % | 8,54 | 59 % |
| 5 | Mistral Large 2 | 13 | 7–19 | 0 | 2,3 | 1,4 | 3,6 | 100 % | 7,32 | 52 % |
Medicine · 9 Items
Ranks 1–2 statistically tied (95 % CI overlaps rank 1)
In Medicine, GPT-5 (top score) contradicts a stored key fact on average 0,2 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. The lowest value is Gemini 2.5 Pro at avg 0,6 per answer. Relative to answer length: 0,81 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 | 85 | 79–90 | +5 | 0,2 | 1,1 | 1,3 | 100 % | 0,81 | 80 % |
| 2 | Claude Opus 4.7 | 81 | 70–89 | -1 | 0,2 | 0,9 | 1,1 | 89 % | 1,09 | 84 % |
| 3 | Mistral Large 2 | 68 | 59–77 | +1 | 0,3 | 0,9 | 1,2 | 100 % | 2,18 | 73 % |
| 4 | Gemini 2.5 Pro | 52 | 45–58 | -2 | 0,2 | 0,4 | 0,6 | 78 % | 2,41 | 69 % |
| 5 | Grok 4.3 | 52 | 47–59 | — | 0,2 | 0,3 | 0,6 | 44 % | 2,59 | 74 % |
Law · 8 Items
In Law, Claude Opus 4.7 (top score) contradicts a stored key fact on average 0,1 times per answer and adds avg 0,6 claims the judge panel flags as false or unsupported; 63 % of answers carry at least one flag. The lowest value is Grok 4.3 at avg 0,1 per answer. Relative to answer length: 0,83 per 1000 output tokens, lowest GPT-5 at 0,60.
| # | 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 | 85 | 80–90 | +8 | 0,1 | 0,6 | 0,7 | 63 % | 0,83 | 92 % |
| 2 | GPT-5 | 68 | 56–79 | +2 | 0,5 | 0,6 | 1,0 | 75 % | 0,60 | 66 % |
| 3 | Gemini 2.5 Pro | 53 | 43–65 | +2 | 0,4 | 0,2 | 0,4 | 25 % | 1,97 | 69 % |
| 4 | Mistral Large 2 | 49 | 41–58 | 0 | 0,4 | 0,6 | 1,0 | 88 % | 3,28 | 65 % |
| 5 | Grok 4.3 | 48 | 35–59 | — | 0,0 | 0,1 | 0,1 | 13 % | 0,65 | 68 % |
Business law · 8 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 0,6 times per answer and adds avg 1,2 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 1,5 per answer. Relative to answer length: 0,69 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 | 79 | 71–87 | -1 | 0,6 | 1,2 | 1,8 | 100 % | 0,69 | 70 % |
| 2 | Claude Opus 4.7 | 77 | 68–84 | +5 | 0,8 | 2,2 | 2,9 | 100 % | 1,54 | 79 % |
| 3 | Gemini 2.5 Pro | 50 | 41–59 | +6 | 0,9 | 0,7 | 1,5 | 88 % | 3,82 | 63 % |
| 4 | Grok 4.3 | 44 | 28–61 | — | 1,5 | 0,6 | 2,1 | 100 % | 5,22 | 61 % |
| 5 | Mistral Large 2 | 43 | 30–55 | -8 | 1,6 | 1,5 | 3,2 | 100 % | 3,83 | 48 % |
Raw data of this run
raw.jsonl and progress.jsonl are the public streams with response texts stripped (contamination protection, methodology §7).