FrontierBioML: Evaluating agents on machine learning tasks in computational biology (FrontierBioML)
We show this table for reference; we do not rank on it.
Valen Labs' benchmark of AI agents doing machine learning work in computational biology: 25 tasks in antibody development, designed binders, protein engineering, structural biology and functional genomics. Each agent builds a model from experimental data in an offline sandbox, and its prediction script is scored on held-out measurements, from 0 (a simple baseline) to 100 (a perfect answer), over three runs per task.
Mean score on FrontierBioML — Valen Labs data of 13 November 2026
We mirror the published mean score view for FrontierBioML. Claude Mythos 5.2 leads the public snapshot at 40.3%, followed by Gemini 4 Carbon (39.0%) and Gemini 4 Argon (36.8%). We do not use these results to rank models overall.
Claude Mythos 5.2
Anthropic
max reasoning
Claude Code · $48.17 and 2.5 h per run
Gemini 4 Carbon
high reasoning
OpenCode · $8.52 (estimated) and 1.8 h per run
Gemini 4 Argon
high reasoning
OpenCode · 1 of 75 runs not graded · $6.71 and 1.6 h per run
11 modelsAgenticCurrentDisplay onlyUpdated Valen Labs data of 13 November 2026
Mean score table (11 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Claude Mythos 5.2Anthropicmax reasoningClaude Code · $48.17 and 2.5 h per run | 40.3% | Not reported | Closed |
| 2 | Gemini 4 CarbonGooglehigh reasoningOpenCode · $8.52 (estimated) and 1.8 h per run | 39.0% | Not reported | Closed |
| 3 | Gemini 4 ArgonGooglehigh reasoningOpenCode · 1 of 75 runs not graded · $6.71 and 1.6 h per run | 36.8% | Not reported | Closed |
| 4 | Claude Opus 5.5Anthropicmax reasoningClaude Code · $12.58 and 1.2 h per run | 36.5% | Not reported | Closed |
| 5 | GPT-6.1 SolOpenAImax reasoningCodex CLI · 17 of 75 runs blocked by a biosafety filter · $1.96 and 26 min per run | 29.1% | Not reported | Closed |
| 6 | GPT-6 AstraOpenAImax reasoningCodex CLI (Azure) · 19 of 75 runs blocked by a biosafety filter · $10.30 and 33 min per run | 28.8% | Not reported | Closed |
| 7 | Muse WonderMetamax reasoningOpenCode · $6.38 and 34 min per run | 26.9% | Not reported | Closed |
| 8 | Kimi K3.1 ProMoonshot AImax reasoningOpenCode · 4 of 75 runs not graded · $2.29 and 34 min per run | 25.8% | Not reported | Open |
| 9 | Grok 4.7xAIxhigh reasoningOpenCode · 1 of 75 runs not graded · $5.04 and 39 min per run | 24.0% | Not reported | Closed |
| 10 | DeepSeek V4.1 Flash 1017DeepSeekmax reasoningOpenCode · 5 of 75 runs not graded · $0.21 and 1.5 h per run | 22.4% | Not reported | Open |
| 11 | Nemotron 3.5 Super VL 120B A12BNVIDIAon reasoningOpenCode (PrimeIntellect) · 13 of 75 runs not graded · $0.11 and 12 min per run | 11.0% | Not reported | Open |
How to read this leaderboard
Higher is better. Claude Mythos 5.2 leads at 40.3, ahead of Gemini 4 Carbon (39.0), Gemini 4 Argon (36.8) and Claude Opus 5.5 (36.5). GPT-6.1 Sol (29.1) and GPT-6 Astra (28.8) lose ground mainly to OpenAI's biology filter, which blocked 17 and 19 of their 75 runs; on the runs that were scored, Valen puts them near the top group. Scores are not comparable across tasks, and on several tasks the leaders are within a point of each other. Mythos 5.2 is also the most expensive, at about $48 per run against $13 for Opus 5.5 and $9 for Gemini 4 Carbon.
Operator receipt: 11 sourced rows are currently displayable on this page; the leading published row is Claude Mythos 5.2 at 40.3%.
Honest limit: Twenty-five tasks is a small set, and three runs per task leave real noise: on one task (CAR binder triage) Valen says gaps under about 9 points are noise. Each row rates a model and harness together, at the reasoning effort Valen chose. OpenAI models ran without verified life-science access, while the Claude models ran under Anthropic's Life Science Verification Program, so refusals weigh on the OpenAI rows. The held-out test data is private; only a public preview of some tasks is released. Valen Labs builds and grades the benchmark itself. BenchLM shows FrontierBioML as display only; it does not enter overall or category rankings.
How FrontierBioML is shown here
We mirror the FrontierBioML leaderboard from Valen Labs. Each agent gets 25 machine learning tasks in computational biology (antibodies, designed binders, protein engineering, structural biology and functional genomics), with experimental data and a brief, in an offline sandbox with no GPU and about six hours per task. It submits a prediction script, and on most tasks a shortlist of candidates, which Valen runs twice on held-out data.
Each task is scored from 0 (a simple baseline) to 100 (a perfect answer). The score is the mean over the 3 runs of each task, averaged over all tasks; a run blocked by a biosafety filter or failing the grading checks counts as 0. Every model runs in a harness (Claude Code, Codex CLI or OpenCode), so a row rates the model and harness together. The task baselines average 13.1. FrontierBioML is display only on BenchLM.
Snapshot
The published FrontierBioML snapshot places Claude Mythos 5.2 first at 40.3%. The third row is 3.5 points behind. The broader top-10 range is 17.9 points, so the table still separates the published systems.
11 models have been evaluated on FrontierBioML. The benchmark falls in the Agentic category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. FrontierBioML is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About FrontierBioML
Year
2026
Tasks
25 machine learning tasks in computational biology across antibodies, designed binders, protein engineering, structural biology and functional genomics
Format
Offline agent sandbox (4 to 32 CPUs, no GPU, about 6 hours); the submitted prediction script runs twice on held-out data; 3 runs per task, 75 per model
Difficulty
Building and validating models on messy biological data that generalize to held-out antigens, families, cell types and targets
Each task gives the agent real experimental data, sometimes deliberately messy (stale identifiers, contamination, swapped labels), and a written brief: rank antibody variants from sequencing counts, predict nanobody binding to unseen antigen variants, score docking poses, infer gene networks from knock-down screens, forecast cancer gene dependencies, choose a panel of enzyme or binder designs under a budget. The agent works offline in a sandbox with 4 to 32 CPUs, no GPU and about six hours. It submits a prediction script and, on most tasks, a shortlist of candidates. Valen runs the script twice in a separate grading sandbox on held-out data; the two runs must give identical output. Each task's metrics are scaled so that its reference baseline sits at or near 0 and a perfect answer at 100 (the baselines average 13.1). The headline is the mean of per-task mean scores over three runs per task, 75 runs per model. A run blocked by a provider's biosafety filter or failing the grading checks counts as 0. Models run in a harness: Claude models in Claude Code, OpenAI models in Codex CLI, the rest in OpenCode. BenchLM recomputes the score from the run-level data on the FrontierBioML site and keeps each task's mean in its history.
Freshness and provenance
Version
FrontierBioML 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does FrontierBioML measure?
Whether an AI agent can do machine learning work in computational biology: build a model from experimental data on antibodies, proteins, structures or gene screens, and produce predictions or a shortlist of candidates that hold up on held-out measurements. Each of 25 tasks is scored from 0 (a simple baseline) to 100 (a perfect answer).
Which model leads FrontierBioML?
Claude Mythos 5.2 in Claude Code leads at 40.3, followed by Gemini 4 Carbon at 39.0, Gemini 4 Argon at 36.8 and Claude Opus 5.5 at 36.5 (Valen Labs data of November 13, 2026).
Why do the GPT-6 models score lower on FrontierBioML?
OpenAI's biology filter blocked 17 of GPT-6.1 Sol's 75 runs and 19 of GPT-6 Astra's, and blocked runs count as 0. Valen was not verified for OpenAI's life-science access program, while the Claude models ran under Anthropic's Life Science Verification Program.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.