Sources
Every score on this site comes from one place: Arcophos's own run of the benchmark harness. No number here is copied from a vendor system card, a launch post, or another leaderboard. The document of record for that run is the methodology page, which states the authoring, audit, and grading protocol the scores were produced under.
The tables below list, per board and per model, the score, its 95 percent confidence interval, the number of released tasks scored, the snapshot month, and the locator: the task set analysis file the number was read from and the date of that snapshot. The site shows 34 scores across 2 boards; none are awaiting source documentation.
Documents
| Arcophos run | Health Optimization Bench: Arcophos harness run |
The run itself: harness analysis over the release set of each task set, one answer per task, tools disabled. Locators name the task set directory in the Arcophos harness repository (tasksets/glp1-ev for the incretin evidence suite, tasksets/mb2ev for the subject suites) and the snapshot date the analysis was read. The public sample and the open-source runner are linked from the data sample section.
Subject suites
| # | model | score | 95% CI | tasks | snapshot | source | |
|---|---|---|---|---|---|---|---|
| 1 | Claude Fable 5 Anthropic | 70.9 | 68.0–73.8 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.709, ci 0.680-0.738, n 257” | |
| 2 | Claude Opus 5 Anthropic | 69.3 | 66.2–72.4 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.693, ci 0.662-0.724, n 257” | |
| 3 | Grok 4.6 SpaceX AI | 66.8 | 63.8–69.8 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.668, ci 0.638-0.698, n 257” | |
| 4 | GPT-5.6 Sol (max) OpenAI max reasoning effort | 66.6 | 63.4–69.6 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.666, ci 0.634-0.696, n 257” | |
| 5 | GPT-5.6 Sol (high) OpenAI high reasoning effort | 64.6 | 61.5–67.6 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.646, ci 0.615-0.676, n 257” | |
| 6 | Kimi K3 Moonshot AI | 59.9 | 56.5–63.3 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.599, ci 0.565-0.633, n 257” | |
| 7 | Muse Spark Meta | 57.2 | 53.9–60.5 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.572, ci 0.539-0.605, n 257” | |
| 8 | Claude Fable 5.1 Anthropic | 47.3 | 42.2–52.2 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.473, ci 0.422-0.522, n 257” | |
| 9 | Gemini 3.6 Google | 39.7 | 36.2–43.1 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.397, ci 0.362-0.431, n 257” | |
| 10 | TM | Inkling Thinking Machines | 35.6 | 32.3–39.0 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.356, ci 0.323-0.390, n 257” |
| 11 | Claude Sonnet 5 Anthropic | 34.6 | 31.6–37.8 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.346, ci 0.316-0.378, n 257” | |
| 12 | GLM 5.2 Zhipu | 20.7 | 18.1–23.4 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.207, ci 0.181-0.234, n 257” | |
| 13 | MiniMax M3 MiniMax | 18.3 | 15.7–20.9 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.183, ci 0.157-0.209, n 257” | |
| 14 | MAI Thinking Microsoft AI | 17.5 | 15.1–20.0 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.175, ci 0.151-0.200, n 257” | |
| 15 | Mistral Medium 3.5 Mistral | 9.2 | 7.4–11.1 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.092, ci 0.074-0.111, n 257” | |
| 16 | Nemotron 3.5 Lightning NVIDIA | 4.9 | 3.6–6.2 | 257 | September 2026 | Arcophos run tasksets/mb2ev/analysis.json, snapshot 2026-09-10 “mean 0.049, ci 0.036-0.062, n 257” | |
Incretin Therapeutics: Evidence Synthesis
| # | model | score | 95% CI | tasks | snapshot | source | |
|---|---|---|---|---|---|---|---|
| 1 | Claude Fable 5.1 Anthropic | 84.9 | 80.1–89.3 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.849, ci 0.801-0.893, n 89” | |
| 2 | Claude Fable 5 Anthropic | 83.8 | 79.1–88.1 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.838, ci 0.791-0.881, n 89” | |
| 3 | GPT-6 Astra (max) OpenAI max reasoning effort | 82.2 | 77.7–86.4 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.822, ci 0.777-0.864, n 89” | |
| 4 | Grok 4.6 SpaceX AI | 81.3 | 77.0–85.3 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.813, ci 0.770-0.853, n 89” | |
| 5 | Claude Opus 5 Anthropic | 78.3 | 72.9–83.4 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.783, ci 0.729-0.834, n 89” | |
| 6 | GPT-5.6 Sol (max) OpenAI max reasoning effort | 77.9 | 73.3–82.2 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.779, ci 0.733-0.822, n 89” | |
| 7 | Kimi K3 Moonshot AI | 77.8 | 73.0–82.4 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.778, ci 0.730-0.824, n 89” | |
| 8 | GPT-5.6 Sol (high) OpenAI high reasoning effort | 77.7 | 73.1–82.0 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.777, ci 0.731-0.820, n 89” | |
| 9 | Muse Spark Meta | 68.4 | 62.8–73.9 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.684, ci 0.628-0.739, n 89” | |
| 10 | Gemini 3.8 Flash Google | 67.8 | 61.7–73.6 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.678, ci 0.617-0.736, n 89” | |
| 11 | Gemini 3.6 Google | 57.9 | 51.2–64.3 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.579, ci 0.512-0.643, n 89” | |
| 12 | TM | Inkling Thinking Machines | 53.9 | 47.7–60.1 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.539, ci 0.477-0.601, n 89” |
| 13 | Claude Sonnet 5 Anthropic | 50.1 | 44.3–56.1 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.501, ci 0.443-0.561, n 89” | |
| 14 | MiniMax M3 MiniMax | 35.5 | 29.6–41.6 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.355, ci 0.296-0.416, n 89” | |
| 15 | MAI Thinking Microsoft AI | 33.0 | 26.8–39.3 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.330, ci 0.268-0.393, n 89” | |
| 16 | GLM 5.2 Zhipu | 29.2 | 23.9–34.9 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.292, ci 0.239-0.349, n 89” | |
| 17 | Mistral Medium 3.5 Mistral | 14.4 | 10.6–18.5 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.144, ci 0.106-0.185, n 89” | |
| 18 | Nemotron 3.5 Lightning NVIDIA | 7.8 | 4.7–11.3 | 89 | September 2026 | Arcophos run tasksets/glp1-ev/analysis.json, snapshot 2026-09-10 “mean 0.078, ci 0.047-0.113, n 89” | |
Protocol in brief
Tasks are authored one at a time by one of five frontier model families, audited for material error and citation fidelity by families that did not write them, and removed only when a second family independently confirms a material verdict. Candidate models answer closed book, one answer per task, tools disabled. Each answer is graded blind in three passes spread across three grader families, with the task's authoring family excluded from its panel; any split criterion escalates to a fourth, uninvolved family. Scores are consensus rubric credit on a 0 to 100 scale. Confidence intervals are nonparametric bootstraps over tasks, 10,000 resamples, fixed seed. Scores are comparable only within an identical task set, grading panel, and pass count. The full protocol, the measured verification outcomes, and the safety scoring rule are on the methodology page.