Model results
Terminal-Bench
Choose a shared evaluation protocol to compare model scores. Versions, sources and configurations stay separate.
Atlas snapshot captured 10/19/2026, 10:20:28 AM UTC. This is an import timestamp, not the evaluation date. Coverage may be incomplete.
Reviewed recommendation and limits
Start with Terminal-Bench 4.0 for terminal-based agent work. Compare the same task release, agent and resource settings.
Terminal-Bench 4.0.0 on Harbor. Preserve agent, model, reasoning effort, task resources, trial count and resolution-rate uncertainty. Earlier 2.0, 2.1 and 3.0 results remain separate.. Imported groups below may use different protocols; this review does not certify each run. Public solutions, reward hacking and task-specific tuning can inflate results. Earlier editions and vendor-reported runs do not establish the current reviewed protocol. Terminal completion is different from repository issue repair.
Terminal-Bench 4.0
- Source
- Anthropic (reviewed publisher results)
- Version
- 4.0
- Measurement date
- 2026-09-22
- Metric
- % · Higher is better
- Evaluation scope
- Vendor-reported in Anthropic's Claude Opus 5.5 launch table; adaptive thinking at max effort unless noted; evaluated with production safeguards on, and where they intervened, cybersecurity tasks fell back to Claude Opus 4.8 and biology/frontier-LLM tasks to Claude Opus 5 (Anthropic says this likely lowers scores); xhigh effort (its highest score); standard error ±2.6 pts; Anthropic's setup reproduces the public leaderboard's Opus 5 result within noise (52.3% vs 51.8%, 5 trials/task, Claude Code harness); dated to the announcement
Positions are withheld because at least one result lacks explicit independent provenance or a documented evaluation scope. Missing metadata is shown as unreported; matching imported metadata is not independent verification of the protocol.
Claude Opus 5.5
anthropic/claude-opus-5-5
66.4%
View modelEvidence model-card-reported