RealmPathology Benchmark: leaderboard
RealmPathology Benchmark (micro1, 2026): expert-authored tasks that give an agent one or more real anatomic pathology report PDFs and ask for "the structured interpretation a tumor board can act on". Unlike pathology boards that grade image reading on whole-slide images, such as MedFlowBench's QuPath tasks, the input is the written report: extraction criteria are met 93% of the time, and the score turns on interpretation, staging, management and not overreaching. Of 1,071 rubric criteria, 299 are penalties; "a penalty fires when a summary invents a finding, misclassifies the case, or recommends beyond what the report supports, and it subtracts from the score." This release is "a new comparison identity", "not continuous with the earlier release", whose rubrics had no penalties. Protocol: a sandboxed OpenCode container with a shell, file tools, a todo list, and web search and web fetch; three rollouts per task; one judge, GPT-5.4 mini (snapshot of 17 March 2026, medium effort), grades every criterion against a hidden expert answer, one call per criterion. Efforts from the methodology note: Claude Fable 5.1, Claude Sonnet 5.5, Claude Opus 5.5 and Claude Opus 5 at max; Muse Wonder at max via OpenRouter; Muse Spark 1.3, Muse Spark 1.1 and Grok 4.7 at xhigh; Grok 4.6 and Gemini 3.8 Flash at high; Claude Sonnet 5 at default; GPT-6 Astra and GPT-6 Sol at max "with the subagent task tool denied". Score: mean reward (%), the mean over tasks of each model's mean reward over three rollouts, shown times 100 (the report prints 0.780 for 78.0); reward is "earned positive weight over total positive weight, floored at zero". Rows carry those efforts; Gemini 4 Argon (high) and Kimi K3.1 Pro were evaluated after the report ("The report below reflects only the original evaluations except the sample task runs.") and are listed as printed. Higher is better. Source: https://www.micro1.ai/benchmark/realm-pathology
Metric: Mean reward (%). Source: www.micro1.ai. Saturation forecast: Around 2030. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5.1 (Max) | 78 | #10 (Claude Fable 5.1) |
| 2 | Claude Sonnet 5.5 (Max) | 77.3 | #14 (Claude Sonnet 5.5) |
| 3 | Claude Opus 5.5 (Max) | 77 | #3 (Claude Opus 5.5) |
| 4 | Muse Wonder (Max) | 76.4 | #18 (Muse Wonder) |
| 5 | Muse Spark 1.3 (xHigh) | 75.8 | #31 (Muse Spark 1.3) |
| 6 | Claude Opus 5 (Max) | 75.2 | #16 (Claude Opus 5) |
| 7 | Muse Spark 1.1 (xHigh) | 74.9 | #55 (Muse Spark 1.1) |
| 8 | Kimi K3.1 Pro | 73.3 | #17 |
| 9 | Grok 4.7 (xHigh) | 73.1 | #26 (Grok 4.7) |
| 10 | Grok 4.6 (High) | 72.1 | #29 (Grok 4.6) |
| 11 | Gemini 4 Argon (High) | 72 | #9 (Gemini 4 Argon) |
| 12 | Claude Sonnet 5 | 71.7 | #65 |
| 13 | GPT-6 Astra (Max) | 70.3 | #6 (GPT-6 Astra) |
| 14 | GPT-6 Sol (Max) | 69.3 | #19 (GPT-6 Sol) |
| 15 | Gemini 3.8 Flash (High) | 66.3 | #30 (Gemini 3.8 Flash) |
No result here: #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=realmpathology-benchmark · How It Works · Data refreshed daily, snapshot 2026-11-05.