Resolution rate
The reported task success percentage for the listed model and agent harness configuration, paired with its published 95% confidence interval.
Benchmarks · Published Frontier Results
See where frontier agents stand and what still stumps them.
Explore official Terminal-Bench leaderboards across versions 4.0, 3.0, and 2.1: the agent harness, model, measured resolution rate, Pass@2, and 95% confidence interval behind each entry.
TB 4.0 · TB 3.0 · TB 2.195% Confidence IntervalsUpstream Harbor Attribution
FRONTIER HEADROOM & COHORT BAR
- Live upstream Harbor syncTB 4.0 / 3.0
Server-fetched scores, Pass@2, and confidence bounds per version.
- Multi-step coupling barPASS@K
Tasks demand genuine domain diagnosis where shallow heuristics fail.
- Zero shortcut verificationSTAGE 15
Stage 15 Abomination tests every check suite against reward hacking.
Compare entries strictly within the same dataset version
Live Harbor feed
Official Terminal-Bench leaderboards.
Scores are fetched server-side from Harbor’s public leaderboard API and attributed to the benchmark maintainers. Select a version below to inspect its published entries.
Fetching published results…
Metric guide
Understand what each column measures.
Headroom at the top of the leaderboard is why Crossing Hurdles exists: well-specified tasks with realistic multi-step coupling separate genuine reasoning from pattern recall.
Pass@2
The probability estimate of solving a task in at least one of two independent attempts when provided by the upstream benchmark release.
Versioned datasets
Each dataset version defines a distinct task distribution and execution horizon. Always compare entries within the same benchmark version.
Build the next frontier gate
Design tasks that challenge top-ranked agents.
Author a frontier-hard Terminal-Bench 4.0 task in our live studio to join Crossing Hurdles.