BioXArena-Pro - Network Biology - Harness Biomni: leaderboard

BioXArena-Pro - Network Biology - Harness Biomni: Native task score x100, failures zero (0-100). BioXArena-Pro release v5.1.1: 64 biomedical ML tasks, 8 domains x8, every score 0-100 and failed task zero, one run/configuration/task, 1 GPU and 7200s/task. Primary CSV retains exact model_dir, scaffold, backend, route, own-run mean, substituted tasks and all domain scores. MLEvolve Opus 5.5 source substitutes two Opus 5 results; its mixed mean/domain panels are held. Other 21 configurations retain separate harnesses and served model identity (DeepSeek V4.1 Flash 1017). Gateway machine/routing differences are retained, never inferred from current defaults. Cost/time/token columns not capability. Runtime effort unreported. Native source condition: Harness Biomni.

Metric: Native task score x100, failures zero (0-100). Source: raw.githubusercontent.com. Saturation forecast: Rough model projection: around 2026. 4 models tracked.

Top models

#ModelScoreOverall rank
1Claude Opus 5.5 (Benchmark Sweep Run: source=bioxarena_pro; harness=Biomni; route=institutional proxy; model_dir=biomni__claude-opus-5-5; effort=unreported; tasks=64; GPU=1; max seconds=7200)42.53
2Gemini 3.8 Flash (Benchmark Sweep Run: source=bioxarena_pro; harness=Biomni; route=OpenRouter [paid]; model_dir=biomni__gemini-3.8-flash; effort=unreported; tasks=64; GPU=1; max seconds=7200)41.49
3GPT-6 Astra (Benchmark Sweep Run: source=bioxarena_pro; harness=Biomni; route=institutional proxy; model_dir=biomni__gpt-6-astra; effort=unreported; tasks=64; GPU=1; max seconds=7200)40.23
4DeepSeek-V4.1-Flash-1017 (Benchmark Sweep Run: source=bioxarena_pro; harness=Biomni; route=OpenRouter [paid]; model_dir=biomni__deepseek-v4.1-flash-1017; effort=unreported; tasks=64; GPU=1; max seconds=7200)28.97

Interactive version: theaggregate.ai/benchmark?slug=bioxarena-pro-network-biology-harness-biomni · How It Works · Data refreshed daily, snapshot 2026-11-12.