Intelligence Infrastructure for Biology
Live Last read · last published · next check
Following yesterday's Grok 4.6 release, we ran it against our short-horizon biology tasks on benchmarks.bio. Across 1716 trajectories we generated, we find that Grok 4.6 sits roughly at Opus 5/GPT-Sol-5.6 level of intelligence while being far cheaper.
A few days ago Mistral released Shieldstral, its new safety classifier, without disclosing its performance on major biosecurity related tasks, such as viral engineering.
An in depth analysis across our benchmark suite
Today we released results for Kimi-K3, an open-source LLM boasting GPT-5.6/Mythos-level coding-benchmark scores, across our short-horizon therapeutics and -omics benchmarks on benchmarks.bio.
A verifiable benchmark for variant discovery, statistical genetics, and personal genomics
In current generation models, behavioral priors introduce by harnesses such as Claude Code, and Codex cause substantial performance swings on our benchmarks.
A verifiable benchmark for practical decisions about taxonomy, variants, AMR, source tracking, anomaly detection, and engineered sequences in pathogen genomic surveillance workflows.
A paired benchmark for capability and caution in agentic biosecurity risk assessment
Access the Latch platform using Claude Code, Codex, or any agent
Welcoming Harmon Bhasin, Evan Seeyave, and John Wang to the team.
A verifiable benchmark for recovering scientific conclusions from raw single-cell data
A verifiable benchmark for practical decisions about potency, mechanism, exposure, safety, and efficacy
Benchmarking frontier models on practical CUT&Tag/CUT&RUN, ATAC-seq, ChIP-seq, and DNA methylation workflows
Two rounds of independent expert attempts define a verified subset of 115 spatial biology tasks and expose ambiguity in benchmark specification and grading.
Evaluating whether AI agents can recover complex scientific conclusions from raw spatial biology data
Benchmarking frontier models on messy, real-world single cell data analysis
Biology will not jump straight to autonomous AI scientists. Like software, it will first accelerate where work is executable, feedback-rich, and economically bottlenecked: data analysis.
Overall accuracy for GPT-5.5 and Opus 4.7 remains flat on SpatialBench. Scientist-reviewed trajectories reveal persistent gaps in assay-aware biological judgment.
394 verifiable problems across six sequencing platforms: 10X Chromium, BD Rhapsody, CSGenetics, Illumina, MissionBio, ParseBio
first contribution towards a broader family of benchmarks spanning major data types // 146 verifiable problems spanning 5 spatial platforms and 7 task categories
