Ranking run 2026-07-15
Claude Opus 4.7 leads with 79/100 across all domains. Tax law: Claude Opus 4.7 (72) · Medicine: GPT-5.6 Sol (85) · Law: Claude Opus 4.7 (81) · Business law: GPT-5 (84).
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 | 20 / 29 | 79 | 76–82 | 0 | 59 | 1,44 | — |
| 2 | GPT-5 | 0 | 32 / 56 | 77 | 74–81 | +2 | 71 | 0,61 | — |
| 2 | GPT-5.6 Sol n=96/100 | 4 (4 T) | 91 / 230 | 77 | 73–81 | 0 | 70 | 0,26 | — |
| 3 | Claude Opus 4.8 | 0 | 18 / 29 | 74 | 71–77 | -1 | 70 | 0,99 | — |
| 4 | Kimi K2.6 n=80/100 | 20 (16 T) | 191 / 318 | 60 | 56–65 | +2 | 52 | 0,29 | — |
| 5 | Grok 4.5 | 0 | 18 / 30 | 58 | 54–62 | +2 | 75 | 3,65 | — |
| 6 | Gemini 3.1 Pro | 0 | 20 / 32 | 54 | 49–58 | -1 | 71 | 4,80 | — |
| 7 | Mistral Large 2 | 0 | 9 / 15 | 50 | 45–54 | -1 | 48 | 4,48 | — |
| 8 | GLM-4.6 | 0 | 54 / 97 | 45 | 40–48 | 0 | 65 | 0,51 | — |
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 · 15 Items
Ranks 1–4 statistically tied (95 % CI overlaps rank 1)
In Tax law, Claude Opus 4.7 (top score) contradicts a stored key fact on average 0,7 times per answer and adds avg 2,0 claims the judge panel flags as false or unsupported; 100 % of answers carry at least one flag. The lowest value is Grok 4.5 at avg 1,3 per answer. Relative to answer length: 2,06 per 1000 output tokens, lowest GPT-5.6 Sol at 0,31.
| # | 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 | 72 | 65–79 | -3 | 0,7 | 2,0 | 2,7 | 100 % | 2,06 | 79 % |
| 2 | GPT-5.6 Sol | 70 | 52–85 | 0 | 0,9 | 1,3 | 2,2 | 100 % | 0,31 | 74 % |
| 3 | Claude Opus 4.8 | 65 | 57–72 | -2 | 0,4 | 1,4 | 1,8 | 100 % | 1,32 | 70 % |
| 4 | GPT-5 | 65 | 55–76 | 0 | 1,1 | 0,8 | 1,9 | 93 % | 0,75 | 66 % |
| 5 | Grok 4.5 | 49 | 43–56 | +2 | 0,8 | 0,5 | 1,3 | 87 % | 4,29 | 57 % |
| 6 | Gemini 3.1 Pro | 44 | 36–51 | -4 | 0,9 | 0,7 | 1,6 | 80 % | 7,18 | 70 % |
| 7 | Kimi K2.6 | 35 | 21–49 | +2 | 1,4 | 2,0 | 3,4 | 100 % | 0,40 | 62 % |
| 8 | Mistral Large 2 | 24 | 15–35 | -4 | 2,3 | 1,7 | 4,0 | 100 % | 8,13 | 48 % |
| 9 | GLM-4.6 | 23 | 16–31 | +2 | 1,6 | 1,2 | 2,8 | 93 % | 0,79 | 62 % |
Medicine · 20 Items
Ranks 1–4 statistically tied (95 % CI overlaps rank 1)
In Medicine, GPT-5.6 Sol (top score) contradicts a stored key fact on average 0,4 times per answer and adds avg 0,5 claims the judge panel flags as false or unsupported; 90 % of answers carry at least one flag. The lowest value is GLM-4.6 at avg 0,3 per answer. Relative to answer length: 0,48 per 1000 output tokens, lowest GLM-4.6 at 0,14.
| # | Model | top100 | 95 % CI | Δ prev. | contradicts key fact | extra false/unsupp. | avg total | answers affected | per 1000 tok. | IRR |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | GPT-5.6 Sol | 85 | 80–89 | +1 | 0,4 | 0,5 | 0,8 | 90 % | 0,48 | 80 % |
| 2 | Claude Opus 4.7 | 84 | 77–91 | -4 | 0,5 | 0,7 | 1,2 | 100 % | 1,12 | 86 % |
| 3 | GPT-5 | 84 | 75–92 | +3 | 0,4 | 0,5 | 0,9 | 90 % | 0,64 | 78 % |
| 4 | Claude Opus 4.8 | 80 | 73–86 | -5 | 0,3 | 0,4 | 0,7 | 75 % | 0,68 | 79 % |
| 5 | Kimi K2.6 | 71 | 63–79 | +2 | 0,4 | 0,6 | 0,9 | 95 % | 0,19 | 71 % |
| 6 | Mistral Large 2 | 66 | 60–71 | -1 | 0,6 | 0,5 | 1,0 | 85 % | 2,00 | 63 % |
| 7 | Grok 4.5 | 64 | 56–73 | -3 | 0,4 | 0,2 | 0,5 | 75 % | 1,94 | 69 % |
| 8 | Gemini 3.1 Pro | 63 | 56–69 | -1 | 0,3 | 0,3 | 0,5 | 60 % | 2,38 | 74 % |
| 9 | GLM-4.6 | 60 | 54–67 | +3 | 0,3 | 0,1 | 0,3 | 60 % | 0,14 | 77 % |
Law · 35 Items
Ranks 1–4 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 0,6 claims the judge panel flags as false or unsupported; 83 % of answers carry at least one flag. The lowest value is Claude Opus 4.8 at avg 0,7 per answer. Relative to answer length: 1,02 per 1000 output tokens, lowest GPT-5.6 Sol at 0,25.
| # | 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 | 81 | 76–85 | +1 | 0,3 | 0,6 | 0,9 | 83 % | 1,02 | 77 % |
| 2 | Claude Opus 4.8 | 81 | 77–85 | +2 | 0,2 | 0,5 | 0,7 | 80 % | 0,71 | 82 % |
| 3 | GPT-5.6 Sol | 75 | 69–80 | 0 | 0,5 | 0,6 | 1,0 | 86 % | 0,25 | 69 % |
| 4 | GPT-5 | 73 | 65–80 | +2 | 0,5 | 0,6 | 1,1 | 86 % | 0,65 | 71 % |
| 5 | Gemini 3.1 Pro | 60 | 52–67 | 0 | 0,6 | 0,3 | 0,8 | 60 % | 4,19 | 63 % |
| 6 | Grok 4.5 | 58 | 50–65 | +6 | 0,7 | 0,3 | 0,9 | 69 % | 5,14 | 66 % |
| 7 | Kimi K2.6 | 58 | 50–66 | 0 | 0,9 | 0,7 | 1,6 | 79 % | 0,27 | 64 % |
| 8 | GLM-4.6 | 45 | 38–52 | -1 | 0,8 | 0,5 | 1,2 | 77 % | 0,53 | 65 % |
| 9 | Mistral Large 2 | 45 | 37–54 | -4 | 1,3 | 0,6 | 1,9 | 77 % | 6,49 | 66 % |
Business law · 30 Items
Ranks 1–3 statistically tied (95 % CI overlaps rank 1)
In Business law, GPT-5 (top score) contradicts a stored key fact on average 0,2 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. Relative to answer length: 0,49 per 1000 output tokens, lowest GPT-5.6 Sol at 0,22.
| # | Model | top100 | 95 % CI | Δ prev. | contradicts key fact | extra false/unsupp. | avg total | answers affected | per 1000 tok. | IRR |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | GPT-5 | 84 | 81–88 | +3 | 0,2 | 0,8 | 1,1 | 100 % | 0,49 | 80 % |
| 2 | Claude Opus 4.7 | 78 | 74–83 | +5 | 0,3 | 1,9 | 2,3 | 97 % | 1,62 | 74 % |
| 3 | GPT-5.6 Sol | 78 | 71–83 | +1 | 0,6 | 0,8 | 1,3 | 93 % | 0,22 | 71 % |
| 4 | Claude Opus 4.8 | 66 | 60–73 | -3 | 0,7 | 1,1 | 1,8 | 100 % | 1,18 | 66 % |
| 5 | Kimi K2.6 | 64 | 60–68 | +3 | 0,6 | 1,9 | 2,5 | 100 % | 0,32 | 58 % |
| 6 | Grok 4.5 | 58 | 50–64 | -1 | 0,7 | 0,6 | 1,3 | 83 % | 3,35 | 62 % |
| 7 | Mistral Large 2 | 57 | 50–64 | +1 | 0,7 | 1,4 | 2,0 | 97 % | 3,27 | 55 % |
| 8 | Gemini 3.1 Pro | 46 | 38–55 | -1 | 1,0 | 0,7 | 1,7 | 90 % | 5,52 | 54 % |
| 9 | GLM-4.6 | 44 | 39–49 | -3 | 0,8 | 0,8 | 1,6 | 90 % | 0,52 | 53 % |
Raw data of this run
raw.jsonl and progress.jsonl are the public streams with response texts stripped (contamination protection, methodology §7).