Claude Opus 5.5 tops TxBench antibody-discovery benchmark at 67.7%, GPT-6 Astra leads oligo tasks
kenbwork · x · 2026-10-20
benchmarks.bio published new TxBench therapeutics benchmark scores, where agents must recover the original biological result from real experimental data to pass.
- Antibody-Discovery: Claude Opus 5.5 takes the top two spots (67.7% via Claude Code, 66.7% via Pi), with Fable 5.1 at 63.3% in third; GPT-6 Astra (56.0%), Sonnet 5.5, and Grok 4.6 trail.
- Oligo-Discovery (113 evaluations): GPT-6 Astra leads (55.5% via Codex, 53.4% via Pi), Claude Opus 5 via Claude Code scores 48.1%, followed by Grok 4.6, Gemini 3.7 Flash, and DeepSeek V4.1 Flash.
- Scores are reported per model-harness pair since the harness substantially changes results, and grading is deterministic against the actual experimental outcome.
More from Models
- qwen3-4b-deslop: 161 examples trained on a MacBook Air cut slop score 121→12.5 with 68% human-range output — lmoroney · 2026-10-20
- Fine-tuned on a MacBook Air, qwen3-4b-deslop cuts AI slop score from 121 to 12.5 per 1,000 words — lmoroney · 2026-10-20
- Taylor expansion argument: a 16-dim 7th-order expansion has ~250k terms, beyond LLM reasoning — avt_im · 2026-10-20
- Latent Space Digest: DeepSeek V4.1 Flash Leads Open Weights, MiniMax and Mercury 3 Land — Latent Space · 2026-10-20
- Why Perplexity beats Parallel on BrowseComp: code execution is the differentiator, argues Nirant Kabra — NirantK · 2026-10-20
- Grok Bot spotted on Rabbit R1, reportedly powered by r1 — jesselyu · 2026-10-20