AI lab · San Francisco

RL environments for frontier AI labs

We build long-horizon tasks from real expert work, each graded automatically in code, for the labs training frontier models.

Target pass rate for today’s best models
5–30%
Of tasks graded automatically, in code
100%
Test tasks ever used for training
0
Domains of expert work
8
Our mission

We build the hardest work AI can’t do yet, and make it verifiable, so frontier models can learn it on the path to AGI.

Research

Frontier models on expert work

Top models now score 97% on older coding benchmarks. On expert work in finance, science and medicine, the best model still fails a third to half of the tasks.

Best published score per benchmarkAccuracy of the top model. The rest of each bar is still unsolved.
  1. SWE-bench VerifiedFor contrast: an older benchmark
  2. Vals Science IndexLife sciencesTop models1. GPT-6.2 Astra68.4%2. GPT-6 Astra65.1%3. Claude Fable 5.263.7%
  3. Terminal-Bench 4.0Software engineeringTop models1. Claude Fable 5.265.2%2. Claude Opus 5.565.2%3. Claude Sonnet 5.564.1%
  4. MedCodeMedicineTop models1. Claude Opus 563.6%2. Gemini 3.1 Pro Preview59.1%3. Gemini 4 Argon58.8%
  5. Tax Agent BenchFinanceTop models1. Gemini 4 Hydrogen52.3%2. Gemini 4 Carbon51.8%3. Claude Fable 5.150.8%

Public leaderboards from Vals AI, read November 17, 2026: SWE-bench Verified, Terminal-Bench 4.0, Tax Agent Bench, Vals Science Index, MedCode. Scores belong to their publishers. PrimeSet benchmarks are in progress.

Products

Environments, benchmarks and custom builds

For labs training and evaluating frontier models on expert work.

  1. 01

    RL environments

    Environments that mirror real expert work, with verifiers that grade every answer in code.

    • Coding, computer use, research and analysis
    • Automatic, evidence-aware grading
    • Fits your existing training stack
    Explore environments
  2. 02

    Benchmarks and evaluations

    Versioned, reproducible evaluations that show where frontier models fall short, and why.

    • Versioned and repeatable
    • Detailed failure breakdowns
    • Held out from training
    Explore benchmarks
  3. 03

    Custom builds

    Environments built around the skills and domains your lab wants to train next.

    • Scoped with your team
    • Written and reviewed by domain experts
    • Delivered with its own sealed test set
    Talk to us
Standards

How we build every environment

The same bar in every domain, from the first task to the last.

  1. 01

    Verifiable

    Automatic verifiers check every answer and the evidence behind it.

  2. 02

    Frontier-hard

    Calibrated so today’s best models are challenged, and can still learn.

  3. 03

    Uncontaminated

    Test sets are held out and never used for training.

  4. 04

    Expert-built

    Designed and reviewed by people who do the work.

Coverage

Domains we cover

Built with the people who do the work in each field.

  • Software engineeringRepositories, tests and tooling
  • MathematicsProofs and quantitative reasoning
  • Life sciencesBiology and chemistry
  • Physical sciencesPhysics and engineering
  • FinanceReports, models and filings
  • LawContracts and case work
  • MedicineRecords and guidelines
  • Computer useBrowsers, apps and documents