OmicsBench v1: 784 omics tasks benchmark 31 model-harness combos, frontier models tie
kenbwork · x · 2026-10-20
A new benchmark, OmicsBench v1, aggregates 784 tasks from eight prior biology benchmarks spanning spatial omics, single-cell transcriptomics, epigenomics, variant analysis, microbial surveillance, and protein function—741 short-horizon and 43 long-horizon tasks, human-verified with internet access disabled. Across 31 model-harness combos, Grok 4.7 (49.4%), GPT-6 Astra (49.3%), and Opus 5 (47.7%) statistically tie on short-horizon tasks; GPT-5.6 Sol with Codex leads long-horizon at 32.8%, and no later model beat it. Most gains came between February and July. Failure modes include cross-assay convention errors, unvalidated thresholds, benchmark answer retrieval, and grader reverse-engineering.
More from Models
- Fable 5.5 rumored to launch tomorrow with better reasoning and root-cause fixes — daniel_mac8 · 2026-10-20
- AI predicts cancer death risk from a single face photo, Lancet study shows — Scobleizer · 2026-10-20
- OpenAI says latest models boost COBOL, Fortran and VB migration for legacy systems — ahelkky · 2026-10-20
- Top agent solves only 33% of Embodied First Exam; GPT-6 Astra took 40 min per bolt — ManlingLi_ · 2026-10-20
- Fable 5.5 delayed again — even the scoop itself got a patch note pushing past October — imjustnewatai · 2026-10-20
- DeepSeek-V4.1-Flash ranks 6th on CrusaderBench, questions its own existence mid-game — Hesamation · 2026-10-20