Terminal-Bench 4.0: September 2026 Coding Agent Leaderboard

Terminal-Bench 4.0 is an open-source benchmark that measures how well an agent can complete long, realistic pieces of work in a sandboxed terminal. Version 4.0 is a fresh set of 66 community-contributed, maintainer-reviewed tasks — none of which appear in Terminal-Bench 2.1. Tasks span software engineering, machine learning, science, operations, security, hardware, and media.

Each task is built around a real deliverable — a working service, a proof, a CAD model, a trained kernel, a forensic report. The median task is estimated at 4 hours of expert work. Each model is run through the full benchmark 3 times (avg@3), each run scored as pass@1, and the displayed score is the mean with error bars showing standard error.

Version 4.0 removed 8 tasks that became saturated, refusal-prone, or leaked public solutions. 20 tasks received revised environments and tighter validation verifiers to avoid false-positive passes. An 8-hour agent timeout applies per task.

Rank	Agent	Model	Effort	Resolution Rate	Tokens	Eval Cost
1	Codex	GPT-6 Astra	max	58.2%	1.5B	$3.3k
2	Claude Code	Fable 5.1	max	57.9%	2.7B	$6.2k
3	Codex	GPT-6 Astra	xhigh	57.9%	—	—
4	Claude Code	Fable 5.1	high	54.5%	—	—
5	Claude Code	Opus 5	xhigh	53.9%	—	—
6	Claude Code	Opus 5	max	51.8%	6.5B	$6.0k
7	Claude Code	GLM-5.3	max	41.8%	—	—
8	Grok Build	Grok 4.7	xhigh	37.6%	—	—
9	Codex	GPT-5.6 Sol	max	37.3%	—	—

27 models evaluated in the September 23, 2026 snapshot. Full leaderboard at benchlm.ai/benchmarks/terminal-bench-4 and tbench.ai.

1. The efficiency divide is the story

GPT-6 Astra and Claude Fable 5.1 are neck-and-neck at 58.2% vs 57.9%, but Astra needed only 1.5B tokens ($3.3k) compared to 2.7B tokens ($6.2k) for Fable 5.1. On a cost-per-solved-task basis, Astra is nearly 2× more efficient. Opus 5 consumed 6.5B tokens ($6.0k) to reach 51.8% — the most token-hungry model in the top tier.

2. GLM-5.3 is the open-weight breakthrough

Z.ai's GLM-5.3 achieved 41.8% inside the Claude Code harness — surpassing GPT-5.6 Sol (37.3%) and Opus 4.8 (23.6%) while running at accessible pricing. This is the strongest open-weight showing on a frontier-difficulty agent benchmark, and it signals that the gap between open and closed models on agentic tasks is narrowing.

3. Grok 4.7 underperforms expectations

Grok 4.7 (xHigh) managed only 37.6% via Grok Build — behind GPT-5.6 Sol and well behind the frontier tier. This is consistent with independent testing showing high token consumption and modest Intelligence Index gains over Grok 4.6.

4. Lowest run cost vs highest capability

The lowest per-run cost was GPT-5.6 Luna at ~$0.3k (using $6/1M API pricing), but it scores far below the frontier. The cheapest frontier-tier model is GPT-6 Astra at $3.3k for the full benchmark — a data point worth noting for teams evaluating agent economics at scale.

• Each score is avg@3: 3 full benchmark runs, pass@1 per task, mean displayed
• Each task graded by its own verifier — inspects final environment state or produced artifact
• A model must pass the verifier's full test suite to score — partial credit is not given
• 8-hour timeout per task; tasks requiring more time score zero
• Results reflect specific agent+model+configuration variants, not generic product names
• Model choice, settings, and execution configuration materially change outcomes

• BenchLM Terminal-Bench 4.0 leaderboard (benchlm.ai/benchmarks/terminal-bench-4, September 23, 2026 snapshot)
• tbench.ai official results
• Vals.ai Terminal-Bench 4.0 page (updated September 16, 2026)
• Morph LLM coding agent leaderboard (updated September 22, 2026)