Updates

BioInspect added nine evaluations: BixBench3, BioStudyBench, LongevityBench, EpiBench, VariantBench, MetagenomicsBench, and three TxBench evaluations covering antibody discovery, preclinical pharmacology, and oligonucleotide discovery. These additions broaden the catalog from knowledge questions to study-scale analysis, aging biology, and experimental decision-making.

We also expanded BioMysteryBench to all 25 models in Vals AI’s current table, added official per-subbenchmark LABBench2 results, refreshed Latch Bio’s model-and-harness configurations, and added OpenAI’s newer LifeSciBench Gold v1 and GeneBench Pro v13 results. All are source-reported results—not new BioInspect benchmark runs.

Metrics and configurations remain separate: BixBench3 measures artifact scores; BioStudyBench separates tool-assisted analysis from prior knowledge and reports standard deviation across repeats; LongevityBench retains proportions and error in years, with the correct ranking direction. Latch’s current V1 and legacy V0 are labeled according to its updated version guide, with earlier source snapshots retained. Missing scores are not treated as failures, and unreported validity rates or confidence intervals are not inferred.

BioInspect added newly published results for Claude Opus 5.5, Claude Sonnet 5.5, and GPT-6.1 Sol across bioinformatics, single-cell and spatial biology, and healthcare evaluations. These are source-reported evaluations from Latch Bio, Vals AI, Anthropic, and OpenAI—not new BioInspect benchmark runs.

BioMysteryBench now includes all three models: Sonnet 5.5 at 81.1%, GPT-6.1 Sol at 79.6%, and Opus 5.5 at 79.3%, using Vals AI’s Terminus 2 evaluation. We also added Opus 5.5 to Terminal-Bench Science’s Life Sciences subset and new scores for scBench, SpatialBench, their Long variants, and HealthBench.

Results retain their source and configuration. Latch V2 scores remain separate from earlier V1 and paper results; HealthBench separates raw and length-adjusted scores, subsets, and grading setups. GPT-6.1’s ProtocolQA Open-Ended result is explicitly labeled as an adaptation, not the original LAB-Bench multiple-choice evaluation.

Jev achieved 77.4% accuracy (387/500) on the official PubMedQA test split, with 100% valid answers. Its 0.25-second median sample execution time was the lowest measured among the nine models in this BioInspect comparison.

Evaluated through Cloudflare using Inspect, Jev selects structured yes/no/maybe choices rather than generating a free-text answer. All nine models use the same test split and scorer. Timings reflect each recorded model-provider configuration, not a controlled comparison of model inference speed.

The full results include confidence intervals, answer validity, latency, and methodology for all 500 questions.

BioInspect added full 500-question PubMedQA results for Kimi K3, Qwen3.6-Plus, Qwen3.8-Max, Nemotron 3 Super, and Nemotron 3 Ultra. Together with Mistral Large, GLM-4.7-Flash, and DeepSeek V4 Flash, the benchmark now compares eight affordable, open-source, or open-weight models under the same evaluation methodology.

Kimi K3 leads this set at 80.6%, followed by Qwen3.6-Plus at 79.4% and Qwen3.8-Max at 78.2%. Each result includes its sample count, 95% confidence interval, answer-format validity, model route, and execution provenance.

Today we are publishing the first BioInspect Benchmark: an independently executed evaluation of AI models on a biology-focused task using PubMedQA, a biomedical question-answering benchmark based on PubMed abstracts. The initial results include Mistral Large, GLM-4.7-Flash, and DeepSeek V4 Flash.

BioInspect began as a catalog for understanding how AI systems are evaluated across biology, and it will continue to add relevant evaluations as more are released. BioInspect Benchmarks extends that work by running selected evaluations directly and publishing the resulting scores alongside their sample counts, confidence intervals, answer-format validity, model routes, and execution methodology. Published results from external sources remain separate, so readers can distinguish source-reported evidence from evaluations executed by BioInspect.

This is the first of many planned runs. The program will focus first on affordable and open-source or open-weight models, making it easier to understand how accessible systems perform on meaningful biology tasks—not only how the largest proprietary models score. Future updates will add more models, broader biological domains, and additional evaluations where the implementation and evidence can be documented clearly. We hope this work will show how researchers can use more accessible models for biological tasks—an opportunity that will become increasingly important as open models grow more capable.

BioInspect added newly reported GPT-5.6 Sol and Opus 5 results across four demanding single-cell and spatial-biology evaluations published by latch.bio. The update covers both focused analysis tasks and longer-horizon research workflows:

• scBench
  GPT-5.6 Sol 62.1% · Opus 5 60.1%
• SpatialBench
  GPT-5.6 Sol 74.5% · Opus 5 71.0%
• scBench-Long
  Opus 5 41.3% · GPT-5.6 Sol 38.1%
• SpatialBench-Long
  GPT-5.6 Sol 38.9% · Opus 5 25.0%

The results broaden BioInspect's view of current frontier performance and show how strongly outcomes can vary with task duration, evaluation design, and agent harness.