Terminal-Bench 3.0 Leaderboard
The next-generation frontier terminal benchmark. Evaluates autonomous agents across 7 domains and complex multi-container environments with strict if-and-only-if verification.
Last updated: August 19, 2026 · 🆕 Claude Opus 5 (42.7%), GPT-5.6 Sol (34.6%), GLM-5.3 (28.3%), Grok 4.6 (26.5%) · Sources: Snorkel AI · BenchLM · TB 2.1 Leaderboard →
#
Model
Prov
Score
±CI
Agent / Harness
Price (Out)
License
Effort
Src
Terminal-Bench 3.0 Score Comparison
Resolution pass@1 rate across verified benchmark tasks on a 0–100 scale.
About Terminal-Bench 3.0: Developed by Harbor, Laude Institute, Snorkel AI, and Turing. Expands past Terminal-Bench 2.1 by incorporating GPU-enabled nodes (up to H100), multi-container topologies, live microservices, and strict if-and-only-if grading verification. Designed with genuine capability gaps targeting sub-30% initial resolve rates for baseline models.
Key Takeaways: Claude Opus 5 leads decisively at 42.7% (mini-SWE-agent, max reasoning), followed by GPT-5.6 Sol at 34.6% and Claude Fable 5 at 34.0% (34.1%). GLM-5.3 delivers the strongest open-weight performance at 28.3% (+23.7 pts over GLM-5.2's 4.6%) at $1.40/$4.40 per 1M tokens.
Key Takeaways: Claude Opus 5 leads decisively at 42.7% (mini-SWE-agent, max reasoning), followed by GPT-5.6 Sol at 34.6% and Claude Fable 5 at 34.0% (34.1%). GLM-5.3 delivers the strongest open-weight performance at 28.3% (+23.7 pts over GLM-5.2's 4.6%) at $1.40/$4.40 per 1M tokens.
📊 See also: Terminal-Bench 2.1 Leaderboard — standard CLI tasks | DeepSWE v1.1 Leaderboard — long-horizon SWE | SWE-bench Pro Leaderboard
Test these agentic models on live infrastructure
20+ LLMs on CodingFleet. Test command-line execution and complex reasoning in real sandboxes.
🚀 Try on CodingFleet →