Catalog

Grouped by the kind of agent it trains. Items in stock ship now; the rest are built to order.

CategoryOfferingBenchmark / capabilityOne unitSamplesStatus
Coding and terminal agents Terminal-Bench tasks Improves: Terminal-Bench 4.0 One Harbor task directory with its verifier and reference solution 3 sample tasks Available now
Multi-turn coding agent trajectories Improves: SWE One session: a JSONL file, one line per API call 5 session excerpts Available now
Computer-use agents Long-horizon computer-use workflows Improves: OSWorld 2.0 One task environment plus one expert trajectory 7 sample tasks Built to order
3D, graphics and games Blender 3D scene building Improves: 3D modeling in Blender One finished .blend scene and its full request-level trajectory 10 samples Built to order
Browser 3D game development trajectories Improves: Long-horizon web and game development One game: source, per-round requests and full trajectory 3 sample games Built to order

Coding and terminal agents

  • Coding tasks on real codebases, graded by hidden tests; tasks in stock are solved by GPT‑5.6 Sol in 1 to 3 of 5 attempts.

    Improves: Terminal-Bench 4.0

    Successful attempts per stock task
    1–⁠3 / 5
    Attempt success rate in stock
    41%
    Agent time limit per task
    ≤ 8 h
    Available now View details
  • Multi-turn Claude Code sessions on real software projects, recorded one API call per line.

    Improves: SWE

    Claude Opus 5 sessions
    90k+
    Kimi K3 sessions
    50k+
    User turns per session
    1–⁠26
    Available now View details

Computer-use agents

  • OSWorld 2.0-style workflows across desktop apps and websites: a scored task environment plus a full expert trajectory.

    Improves: OSWorld 2.0

    Model turns per trajectory
    245
    Tool calls per trajectory
    1,127
    Checkpoints per task
    30
    Built to order View details

3D, graphics and games

  • Client-style 3D jobs done by an agent in live Blender: the finished .blend scene and its full trajectory.

    Improves: 3D modeling in Blender

    API requests per trajectory
    38–⁠157
    Automatic checks passed
    83 / 83
    Median run time
    51 min
    Built to order View details
  • Playable 3D browser games a coding agent writes from scratch over several rounds of requests, with the full development trajectory.

    Improves: Long-horizon web and game development

    Output tokens per game
    0.68–⁠0.92 M
    Model calls per game
    524–⁠739
    Lines of JavaScript per game
    7.1k–10.9k
    Built to order View details

Tell us what your model needs to get better at.

We reply with samples and a specification.