Frontier models hit 73.3% on SpatialBench 10 months after release, near 75% ceiling
kenbwork · x · 2026-10-28
The SpatialBench authors reflect: 10 months after release, frontier models are approaching 75% accuracy, with GPT-6.1 Terra + Pi setting a new high at 73.3%.
They recount the design: turning real-world spatial transcriptomics analysis (imaging-based methods like MERFISH/Xenium/CosMx vs sequencing-based Visium/Slide-seq pose different challenges) into verifiable problems, with a chunked step structure since errors compound across end-to-end workflows—normalization, clustering, and cell typing each need focused treatment.
At release, biotech was largely unfamiliar with pointing coding agents at experimental data; agents are now becoming reliable at measurement-specific bioinformatics, with distinct failure modes across model families, and harnesses matter.
More from Models
- LiquidAI's LFM3-1.7B edge text-generation model trends on Hugging Face — LiquidAI · 2026-10-28
- Google AI Overviews cites pages that don't contain the cited info, SEO vet flags — lilyraynyc · 2026-10-28
- Leak: Astra 6.1 may launch next week at $30/1M output tokens, claimed to beat Opus 5.5 — bindureddy · 2026-10-28
- Codex Pro user down to 20% usage, publicly begs OpenAI for a reset — AIandDesign · 2026-10-28
- Astra has regressed to 'Opus 5.0 levels of stupid,' frustrated user reports — Because_Bot_Fed · 2026-10-28
- Screenshot Appears to Show GPT-6 and GPT-5.6 Sol in ChatGPT Model Picker — ShelterCultural5725 · 2026-10-28