Terminal-Bench 3.0 at a glance
The new version spans seven domains with richer environments, and the best launch result solved about a third of tasks.
What changed with Terminal-Bench 3.0
Terminal-Bench 3.0, posted on July 30, 2026, builds on version 2.1 with 74 tasks across seven domains. Tasks run longer and use richer setups, including multi-container networks, GPUs, and artifacts such as database snapshots, ML checkpoints, and CAD files. Agent and grader containers are now separate, so results can be re-scored when graders improve.
Terminal-Bench 3.0 in numbers
Values are copied from the linked source on the date shown. They are not IA Compa measurements.
- Best pass rate at launch (GPT-5.6 Sol, Codex)as of · Source for Best pass rate at launch (GPT-5.6 Sol, Codex): tbench.ai (opens in a new tab)
- 34.4%
- Tasks in first releaseas of · Source for Tasks in first release: tbench.ai (opens in a new tab)
- 74
Why Terminal-Bench 3.0 matters for businesses
At launch, the best pass rate was 34.4% for GPT-5.6 Sol running in Codex, followed by Fable 5 in Claude Code at 33.8%. For businesses, the low scores are a reminder that long, unsupervised technical jobs still fail often, and that human review belongs in any agent workflow.
Sources for this story
- Terminal-Bench 3.0 (opens in a new tab)Official
Terminal-Bench · tbench.ai
- Terminal-Bench news (opens in a new tab)Official
Terminal-Bench · tbench.ai
Last verified: . IA Compa summarizes public announcements for business readers. We are not affiliated with the companies mentioned. Always confirm details, pricing, and availability with the official source.
- #benchmarks
- #terminal-bench
- #agents
- #coding
