Ranking run 2026-06-18
Claude Opus 4.7 leads with 80/100 across all domains. Tax law: Claude Opus 4.7 (70) · Medicine: Claude Opus 4.7 (85) · Law: Claude Opus 4.7 (84) · Business law: GPT-5 (80).
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 / 29 | 80 | 77–83 | +1 | 60 | 1,49 | — |
| 2 | GPT-5 | 0 | 24 / 38 | 76 | 72–79 | +2 | 69 | 0,63 | — |
| 3 | Claude Opus 4.8 | 0 | 18 / 28 | 74 | 70–78 | 0 | 69 | 1,02 | — |
| 4 | Kimi K2.6 | 0 | 186 / 772 | 56 | 51–61 | — | 45 | 0,30 | — |
| 5 | Gemini 2.5 Pro | 0 | 21 / 26 | 53 | 49–57 | -1 | 79 | 3,66 | — |
| 6 | Mistral Large 2 | 0 | 10 / 70 | 51 | 45–55 | +2 | 47 | 4,60 | — |
| 7 | Grok 4.3 | 0 | 7 / 8 | 44 | 39–49 | -4 | 68 | 6,30 | — |
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 · 15 Items
Ranks 1–3 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 1,8 claims the judge panel flags as false or unsupported; 93 % 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: 2,06 per 1000 output tokens, lowest Kimi K2.6 at 0,37.
| # | 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 | 70 | 64–77 | -5 | 0,7 | 1,8 | 2,5 | 93 % | 2,06 | 71 % |
| 2 | Claude Opus 4.8 | 69 | 63–74 | +3 | 0,8 | 1,2 | 2,0 | 100 % | 1,42 | 76 % |
| 3 | GPT-5 | 62 | 49–73 | +3 | 1,1 | 1,1 | 2,2 | 100 % | 0,85 | 65 % |
| 4 | Kimi K2.6 | 37 | 26–48 | — | 1,5 | 1,8 | 3,3 | 100 % | 0,37 | 60 % |
| 5 | Gemini 2.5 Pro | 35 | 25–46 | -3 | 1,3 | 0,7 | 2,0 | 87 % | 9,18 | 60 % |
| 6 | Mistral Large 2 | 28 | 18–38 | -2 | 1,7 | 1,3 | 3,0 | 100 % | 6,44 | 60 % |
| 7 | Grok 4.3 | 23 | 14–32 | -3 | 1,7 | 0,7 | 2,4 | 100 % | 10,36 | 51 % |
Medicine · 20 Items
Ranks 1–3 statistically tied (95 % CI overlaps rank 1)
In Medicine, Claude Opus 4.7 (top score) contradicts a stored key fact on average 0,4 times per answer and adds avg 0,8 claims the judge panel flags as false or unsupported; 95 % of answers carry at least one flag. The lowest value is Grok 4.3 at avg 0,4 per answer. Relative to answer length: 1,12 per 1000 output tokens, lowest Kimi K2.6 at 0,18.
| # | 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 | 79–90 | 0 | 0,4 | 0,8 | 1,1 | 95 % | 1,12 | 88 % |
| 2 | GPT-5 | 84 | 77–89 | +4 | 0,5 | 0,6 | 1,0 | 90 % | 0,77 | 78 % |
| 3 | Claude Opus 4.8 | 79 | 71–86 | 0 | 0,3 | 0,3 | 0,6 | 75 % | 0,56 | 76 % |
| 4 | Mistral Large 2 | 68 | 61–74 | -1 | 0,4 | 0,8 | 1,2 | 95 % | 2,28 | 75 % |
| 5 | Kimi K2.6 | 66 | 59–74 | — | 0,5 | 0,5 | 1,0 | 95 % | 0,18 | 62 % |
| 6 | Gemini 2.5 Pro | 59 | 52–66 | -2 | 0,2 | 0,4 | 0,5 | 55 % | 2,31 | 74 % |
| 7 | Grok 4.3 | 53 | 46–60 | +1 | 0,3 | 0,1 | 0,4 | 40 % | 1,79 | 68 % |
Law · 35 Items
Ranks 1–3 statistically tied (95 % CI overlaps rank 1)
In Law, Claude Opus 4.7 (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; 77 % of answers carry at least one flag. The lowest value is Gemini 2.5 Pro at avg 0,3 per answer. Relative to answer length: 1,10 per 1000 output tokens, lowest Kimi K2.6 at 0,26.
| # | 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 | 84 | 80–88 | +4 | 0,2 | 0,8 | 0,9 | 77 % | 1,10 | 84 % |
| 2 | Claude Opus 4.8 | 78 | 72–83 | -2 | 0,3 | 0,5 | 0,8 | 71 % | 0,88 | 77 % |
| 3 | GPT-5 | 74 | 66–80 | +1 | 0,4 | 0,6 | 1,0 | 89 % | 0,56 | 76 % |
| 4 | Gemini 2.5 Pro | 61 | 55–67 | +2 | 0,2 | 0,1 | 0,3 | 29 % | 1,40 | 74 % |
| 5 | Kimi K2.6 | 57 | 49–65 | — | 1,0 | 0,9 | 1,8 | 83 % | 0,26 | 68 % |
| 6 | Mistral Large 2 | 44 | 36–52 | +2 | 1,3 | 0,9 | 2,1 | 83 % | 7,07 | 60 % |
| 7 | Grok 4.3 | 42 | 34–49 | -9 | 1,1 | 0,3 | 1,2 | 66 % | 10,70 | 68 % |
Business law · 30 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 0,7 claims the judge panel flags as false or unsupported; 97 % of answers carry at least one flag. Relative to answer length: 0,51 per 1000 output tokens, lowest Kimi K2.6 at 0,37.
| # | Model | top100 | 95 % CI | Δ prev. | contradicts key fact | extra false/unsupp. | avg total | answers affected | per 1000 tok. | IRR |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | GPT-5 | 80 | 75–85 | 0 | 0,5 | 0,7 | 1,2 | 97 % | 0,51 | 64 % |
| 2 | Claude Opus 4.7 | 78 | 73–83 | +1 | 0,4 | 1,9 | 2,3 | 100 % | 1,69 | 70 % |
| 3 | Claude Opus 4.8 | 69 | 63–75 | +1 | 0,6 | 1,2 | 1,8 | 100 % | 1,13 | 70 % |
| 4 | Kimi K2.6 | 58 | 50–66 | — | 0,8 | 2,0 | 2,9 | 100 % | 0,37 | 53 % |
| 5 | Mistral Large 2 | 58 | 51–64 | +4 | 0,8 | 1,5 | 2,3 | 93 % | 3,76 | 55 % |
| 6 | Grok 4.3 | 51 | 43–59 | -2 | 0,8 | 0,6 | 1,4 | 77 % | 4,67 | 56 % |
| 7 | Gemini 2.5 Pro | 48 | 41–55 | -4 | 0,7 | 0,5 | 1,2 | 87 % | 4,07 | 55 % |
Raw data of this run
raw.jsonl and progress.jsonl are the public streams with response texts stripped (contamination protection, methodology §7).