Terminal-Bench 4.0
The current Terminal-Bench release measures difficult computer work after recalibrating task resources, fixing unstable tasks, and removing tasks that no longer separate frontier systems.
Previous release
Terminal-Bench 3.0 remains available as a historical snapshot. Version 4.0 changes the task set and resource limits, so movement between the two tables is not a like-for-like model comparison.
View Terminal-Bench 3.0How to read this leaderboard
Operator receipt: 10 sourced rows are currently displayable on this page; the leading published row is Claude Opus 5 at 51.82%.
Honest limit: The table compares complete model-and-agent systems, not isolated base models. Version 4.0 also changes resources and the task set, so score changes from older releases mix system progress with protocol changes.
How we use Terminal-Bench 4.0
We mirror the Terminal-Bench 4.0 table across 66 professional computer-work tasks, with 5 trials per task and an 8-hour agent timeout. Version 4.0 fixes 19 tasks and removes 8 tasks that were saturated, refusal-prone, publicly solved, or still affected by quality and platform issues.
Each row measures a model and agent harness together, including Claude Code, Codex, or Grok Build. We keep the raw table display-only. Its normalized result replaces 3.0 inside the same external agentic benchmark family, so the two versions never count as independent evidence.
Snapshot
Tasks completed on Terminal-Bench 4.0 — August 29, 2026 snapshot
We mirror the published tasks completed view for Terminal-Bench 4.0. Claude Opus 5 leads the public snapshot at 51.82%, followed by Claude Fable 5 (44.55%) and GLM-5.3 (41.82%). We do not weight the raw table directly. A normalized version can contribute through our external agentic consensus gate.
Claude Opus 5
Anthropic
Claude Code · max reasoning
Claude Fable 5
Anthropic
Claude Code · max reasoning
GLM-5.3
Z.AI
Claude Code · max reasoning
Tasks completed table (10 models)
ScoreThe published Terminal-Bench 4.0 snapshot places Claude Opus 5 first at 51.82%. The third row is 10.00 points behind. The broader top-10 range is 39.40 points, so the table still separates the published systems.
10 models have been evaluated on Terminal-Bench 4.0. The benchmark falls in the Agentic category. This category carries a 22% weight in BenchLM.ai's overall scoring system. Terminal-Bench 4.0 is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About Terminal-Bench 4.0
Year
2026
Tasks
66 professional computer-work tasks
Format
Task completion rate across 5 trials per task
Difficulty
Frontier autonomous knowledge work
Version 4.0 contains 66 tasks, gives every agent up to eight hours, fixes 19 tasks, and removes eight tasks for saturation, refusals, public solutions, or unresolved quality and platform issues. Those breaking changes required fresh trials, so 4.0 scores are not directly comparable with 3.0 or 2.x. Each row still combines a model with an agent harness; the raw table remains display-only.
BenchLM freshness & provenance
Version
Terminal-Bench 4.0.0
Refresh cadence
Continuous
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
FAQ
What does Terminal-Bench 4.0 measure?
The current Terminal-Bench release measures difficult computer work after recalibrating task resources, fixing unstable tasks, and removing tasks that no longer separate frontier systems.
Which model leads the published Terminal-Bench 4.0 snapshot?
Claude Opus 5 currently leads the published Terminal-Bench 4.0 snapshot with 51.82% tasks completed. BenchLM keeps the raw table display-only. A normalized version can contribute through the external agentic consensus gate, which still requires independent corroboration.
How many models are evaluated on Terminal-Bench 4.0?
The August 29, 2026 snapshot contains 10 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 2,000+ readers.
One email each week. Unsubscribe anytime.