Benchmark · agentic

Terminal-Bench 2

15 results 12 models

Harbor/Terminus-2 89-task terminal-agent suite; 2.0 and 2.1 are the same task set, so one series. File here only when the quote says 2.0 / 2.1 / v2 — not the original suite (terminal-bench), and not the harder 'Terminal-Bench Hard' subset (not in this catalog).

GPT-5.1-Codex-Max Step 3.5 Flash MiniMax M2.7 GPT-5.5 Qwen3.6-27B Tmax Tmax-27B GLM 5.2 FP8 with FP8 KV LLM-as-a-Verifier Grok 4.5 hermes Gemini 3.5 Flash-Lite

Timeline

Date	Model	Score	Source
2026-07-21	Gemini 3.5 Flash-Lite	54.0%	Google releases Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber for agentic workloads
2026-07-21	Gemini 3.5 Flash-Lite	54.0%	Google releases Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
2026-07-08	hermes	55.0%	User benchmarks Mimo v2.5 against DeepSeek V4 Flash and Hermes
2026-07-08	Grok 4.5	83.3%	xAI releases Grok 4.5 with low cost and mixed benchmark performance
2026-07-07	LLM-as-a-Verifier	86.5%	LLM-as-a-Verifier introduces general-purpose verification framework with continuous scoring
2026-07-05	GLM 5.2 FP8 with FP8 KV	79.8%	GLM 5.2 FP8 with FP8 KV achieves 79.8% on Terminal-Bench 2.1
2026-06-23	Tmax-27B	43.0%	Tmax-27B Terminal Agent for Small GPUs with DPPO Training
2026-06-23	Tmax	27.0%	Tmax: A Simple RL Recipe for Terminal Agents
2026-04-25	Qwen3.6-27B	59.3%	Alibaba releases Qwen3.6-27B, beating larger predecessor on coding benchmarks
2026-04-25	GPT-5.5	82.7%	OpenAI releases GPT-5.5, a ground-up retrained model with native omnimodal processing
2026-04-23	GPT-5.5	82.7%	OpenAI releases GPT-5.5 with agentic capabilities and doubled API pricing
2026-04-23	GPT-5.5	82.7%	OpenAI releases GPT-5.5 and GPT-5.5 Pro in API
2026-04-12	MiniMax M2.7	57.0%	MiniMax open-sources M2.7, a self-evolving MoE model scoring 56.22% on SWE-Pro
2026-02-10	Step 3.5 Flash	51.0%	Step 3.5 Flash: open 11B-active MoE model achieves frontier-level agentic performance
2025-11-19	GPT-5.1-Codex-Max	58.1%	OpenAI makes GPT-5.1-Codex-Max the default in Codex CLI