Ranking run 2026-06-21
GPT-5.5 Pro leads with 80/100 across all domains. Tax law: Claude Opus 4.7 (74) · Medicine: GPT-5.5 Pro (88) · Law: Claude Opus 4.7 (82) · Business law: GPT-5.5 Pro (82).
Overall across domains
| # | Model | top100 | 95 % CI | Δ prev. | failures | sec median/p90 | hallu band | per 1000 tok. | Δ own family |
|---|---|---|---|---|---|---|---|---|---|
| 1 | GPT-5.5 Pro n=97/100 | 3 | 227 / 1547 | 80 | 77–83 | — | 75 | 0,10 | — |
| 2 | Claude Opus 4.7 | 0 | 19 / 28 | 78 | 75–81 | -1 | 60 | 1,48 | — |
| 3 | Claude Opus 4.8 | 0 | 21 / 31 | 75 | 72–78 | +1 | 74 | 0,87 | — |
| 3 | GPT-5 | 0 | 22 / 35 | 75 | 71–79 | 0 | 69 | 0,61 | — |
| 4 | Gemini 3.1 Pro | 0 | 22 / 33 | 57 | 53–60 | — | 78 | 3,68 | — |
| 4 | Kimi K2.6 n=98/100 | 2 | 71 / 139 | 57 | 52–61 | — | 46 | 0,27 | — |
| 5 | Mistral Large 2 | 0 | 10 / 69 | 50 | 45–55 | -1 | 50 | 4,35 | — |
| 6 | Grok 4.3 | 0 | 5 / 7 | 46 | 41–50 | -1 | 73 | 5,44 | — |
| 7 | GLM-4.6 | 0 | 42 / 74 | 45 | 40–49 | — | 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 0,5 times per answer and adds avg 1,8 claims the judge panel flags as false or unsupported; 100 % of answers carry at least one flag. The lowest value is Gemini 3.1 Pro at avg 0,9 per answer. Relative to answer length: 1,89 per 1000 output tokens, lowest GPT-5.5 Pro at 0,13.
| # | 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 | 67–80 | +3 | 0,5 | 1,8 | 2,3 | 100 % | 1,89 | 74 % |
| 2 | Claude Opus 4.8 | 67 | 62–73 | -2 | 0,2 | 1,0 | 1,2 | 100 % | 0,88 | 71 % |
| 3 | GPT-5.5 Pro | 67 | 52–80 | — | 1,0 | 1,0 | 1,9 | 92 % | 0,13 | 71 % |
| 4 | GPT-5 | 56 | 39–71 | -7 | 1,3 | 1,0 | 2,3 | 93 % | 0,85 | 73 % |
| 5 | Gemini 3.1 Pro | 54 | 47–61 | — | 0,6 | 0,4 | 0,9 | 80 % | 3,73 | 71 % |
| 6 | Kimi K2.6 | 30 | 18–41 | — | 2,1 | 2,1 | 4,2 | 100 % | 0,42 | 57 % |
| 7 | Mistral Large 2 | 27 | 19–37 | -1 | 2,1 | 1,5 | 3,5 | 93 % | 7,65 | 52 % |
| 8 | GLM-4.6 | 19 | 10–28 | — | 2,1 | 0,8 | 2,7 | 100 % | 0,89 | 58 % |
| 9 | Grok 4.3 | 18 | 10–28 | -9 | 2,1 | 0,7 | 2,8 | 100 % | 12,74 | 68 % |
Medicine · 20 Items
Ranks 1–4 statistically tied (95 % CI overlaps rank 1)
In Medicine, GPT-5.5 Pro (top score) contradicts a stored key fact on average 0,5 times per answer and adds avg 0,4 claims the judge panel flags as false or unsupported; 85 % of answers carry at least one flag. The lowest value is Gemini 3.1 Pro at avg 0,4 per answer. Relative to answer length: 0,18 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-5.5 Pro | 88 | 83–92 | — | 0,5 | 0,4 | 0,8 | 85 % | 0,18 | 81 % |
| 2 | GPT-5 | 83 | 77–89 | +2 | 0,5 | 0,6 | 1,0 | 90 % | 0,72 | 72 % |
| 3 | Claude Opus 4.7 | 81 | 73–88 | -3 | 0,7 | 0,8 | 1,4 | 100 % | 1,37 | 78 % |
| 4 | Claude Opus 4.8 | 81 | 75–86 | +2 | 0,3 | 0,5 | 0,7 | 90 % | 0,66 | 76 % |
| 5 | Kimi K2.6 | 73 | 64–80 | — | 0,5 | 0,7 | 1,1 | 100 % | 0,22 | 74 % |
| 6 | Mistral Large 2 | 68 | 62–75 | +1 | 0,5 | 0,5 | 1,0 | 90 % | 1,91 | 62 % |
| 7 | Gemini 3.1 Pro | 59 | 53–66 | — | 0,3 | 0,2 | 0,4 | 60 % | 1,90 | 72 % |
| 8 | GLM-4.6 | 57 | 51–63 | — | 0,2 | 0,5 | 0,6 | 75 % | 0,25 | 67 % |
| 9 | Grok 4.3 | 55 | 48–61 | 0 | 0,4 | 0,2 | 0,5 | 75 % | 2,39 | 76 % |
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 0,2 times per answer and adds avg 0,6 claims the judge panel flags as false or unsupported; 71 % of answers carry at least one flag. The lowest value is Grok 4.3 at avg 0,6 per answer. Relative to answer length: 0,98 per 1000 output tokens, lowest GPT-5.5 Pro at 0,07.
| # | 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 | 82 | 77–86 | 0 | 0,2 | 0,6 | 0,8 | 71 % | 0,98 | 82 % |
| 2 | GPT-5.5 Pro | 80 | 75–84 | — | 0,3 | 0,5 | 0,7 | 77 % | 0,07 | 76 % |
| 3 | Claude Opus 4.8 | 79 | 73–85 | 0 | 0,2 | 0,4 | 0,6 | 74 % | 0,68 | 79 % |
| 4 | GPT-5 | 78 | 72–83 | +5 | 0,2 | 0,6 | 0,8 | 80 % | 0,46 | 74 % |
| 5 | Gemini 3.1 Pro | 61 | 54–68 | — | 0,4 | 0,2 | 0,6 | 49 % | 3,13 | 73 % |
| 6 | Kimi K2.6 | 56 | 49–63 | — | 0,8 | 0,8 | 1,6 | 80 % | 0,20 | 63 % |
| 7 | Grok 4.3 | 50 | 44–56 | +2 | 0,4 | 0,2 | 0,6 | 46 % | 4,41 | 68 % |
| 8 | GLM-4.6 | 45 | 37–52 | — | 1,1 | 0,3 | 1,1 | 69 % | 0,46 | 64 % |
| 9 | Mistral Large 2 | 44 | 35–53 | -2 | 1,2 | 0,7 | 1,8 | 77 % | 6,17 | 67 % |
Business law · 30 Items
Ranks 1–3 statistically tied (95 % CI overlaps rank 1)
In Business law, GPT-5.5 Pro (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; 100 % of answers carry at least one flag. Relative to answer length: 0,11 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-5.5 Pro | 82 | 77–86 | — | 0,4 | 0,8 | 1,2 | 100 % | 0,11 | 75 % |
| 2 | GPT-5 | 76 | 68–84 | -4 | 0,5 | 0,8 | 1,3 | 93 % | 0,57 | 68 % |
| 3 | Claude Opus 4.7 | 74 | 69–80 | -3 | 0,5 | 1,9 | 2,3 | 100 % | 1,69 | 69 % |
| 4 | Claude Opus 4.8 | 69 | 63–75 | +1 | 0,5 | 1,2 | 1,6 | 97 % | 1,11 | 67 % |
| 5 | Kimi K2.6 | 59 | 52–65 | — | 0,8 | 1,9 | 2,7 | 100 % | 0,30 | 57 % |
| 6 | Mistral Large 2 | 56 | 49–63 | -3 | 0,8 | 1,3 | 2,1 | 93 % | 3,42 | 51 % |
| 7 | Gemini 3.1 Pro | 51 | 44–58 | — | 0,9 | 0,7 | 1,5 | 87 % | 4,85 | 62 % |
| 8 | GLM-4.6 | 49 | 44–56 | — | 0,5 | 1,1 | 1,6 | 97 % | 0,50 | 59 % |
| 9 | Grok 4.3 | 49 | 42–56 | -1 | 0,5 | 0,7 | 1,2 | 87 % | 4,51 | 57 % |
Raw data of this run
raw.jsonl and progress.jsonl are the public streams with response texts stripped (contamination protection, methodology §7).