Terminal-Bench Hard

Agentic

Source ↗ Higher is better

Rank	Model	Provider	Score
1	GPT-5.6 Sol (max)	OpenAI	65.9%
2	Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)	Anthropic	62.9%
3	GPT-5.6 Sol (medium)	OpenAI	62.9%
4	GPT-5.6 Terra (xhigh)	OpenAI	62.9%
5	GPT-5.6 Sol (high)	OpenAI	62.1%
6	GPT-5.6 Sol (xhigh)	OpenAI	61.4%
7	GPT-5.6 Sol (low)	OpenAI	60.6%
8	GPT-5.5 (xhigh)	OpenAI	60.6%
9	GPT-5.5 (high)	OpenAI	59.8%
10	Claude Opus 4.8 (Adaptive Reasoning, Max Effort)	Anthropic	58.3%
11	GPT-5.4 (xhigh)	OpenAI	57.6%
12	GPT-5.6 Terra (max)	OpenAI	57.6%
13	GPT-5.6 Terra (high)	OpenAI	57.6%
14	GPT-5.5 (medium)	OpenAI	57.6%
15	Claude Opus 4.7 (Non-reasoning, High Effort)	Anthropic	54.5%
16	Google: Gemini 3.1 Pro Preview	Google	53.8%
17	Claude Sonnet 4.6 (Adaptive Reasoning, Max Effort)	Anthropic	53.0%
18	GPT-5.3 Codex (xhigh)	OpenAI	53.0%
19	GPT-5.5 (low)	OpenAI	52.3%
20	GPT-5.4 mini (xhigh)	OpenAI	52.3%
21	Claude Opus 4.7 (Adaptive Reasoning, Max Effort)	Anthropic	51.5%
22	GLM-5.2 (max)	Z.ai	50.8%
23	Qwen: Qwen3.7 Max	Alibaba	50.8%
24	KAT Coder Pro V2	Kwaipilot	49.2%
25	GPT-5.5 (Non-reasoning)	OpenAI	49.2%
26	Claude Opus 4.6 (Non-reasoning, High Effort)	Anthropic	48.5%
27	GPT-5.2 (xhigh)	OpenAI	47.0%
28	Claude Opus 4.5 (Reasoning)	Anthropic	47.0%
29	Qwen: Qwen3.7 Plus	Alibaba	47.0%
30	DeepSeek V4 Pro 0424 (Reasoning, Max Effort)	DeepSeek	46.2%
31	Claude Opus 4.6 (Adaptive Reasoning, Max Effort)	Anthropic	46.2%
32	Gemini 3.5 Flash (minimal)	Google	46.2%
33	Claude Sonnet 4.6 (Non-reasoning, High Effort)	Anthropic	46.2%
34	GPT-5.1 (high)	OpenAI	45.5%
35	Muse Spark	Meta	45.5%
36	MoonshotAI: Kimi K2.7 Code	Moonshot AI	44.7%
37	Qwen: Qwen3.6 Plus	Alibaba	43.9%
38	Kimi K2.6 (Reasoning)	Moonshot AI	43.9%
39	Qwen: Qwen3.6 Max PreviewHistorical	Alibaba	43.9%
40	GPT-5.6 Terra (low)	OpenAI	43.9%
41	GLM-5.1 (Reasoning)	Z.ai	43.2%
42	MiMo-V2.5-Pro (Reasoning)	Xiaomi	43.2%
43	GPT-5.4 (low)	OpenAI	43.2%
44	GLM-5 (Reasoning)	Z.ai	43.2%
45	GPT-5.2 (medium)	OpenAI	43.2%
46	GPT-5.5 Instant (May 2026)	OpenAI	42.4%
47	GPT-5.4 nano (xhigh)	OpenAI	42.4%
48	MiniMax: MiniMax M3	MiniMax	42.4%
49	Claude Sonnet 4.6 (Non-reasoning, Low Effort)	Anthropic	42.4%
50	DeepSeek V4 Pro 0424 (Reasoning, High Effort)	DeepSeek	41.7%

Access this data

JSON GET /api/v1

[
  {
    "rank": 1,
    "model": "GPT-5.6 Sol (max)",
    "provider": "OpenAI",
    "score": "65.9090909090909"
  },
  {
    "rank": 2,
    "model": "Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)",
    "provider": "Anthropic",
    "score": "62.8787878787879"
  },
  {
    "rank": 3,
    "model": "GPT-5.6 Sol (medium)",
    "provider": "OpenAI",
    "score": "62.8787878787879"
  }
]

Curl

$ curl -s https://rightrank.ai/api/v1/rankings/terminal-bench-hard.json | jq '.[:3]'

CLI brew install rightrank

$ rightrank top terminal-bench-hard --limit 5
$ rightrank compare claude-opus-4.6 gpt-5.4-pro --category terminal-bench-hard
$ rightrank value terminal-bench-hard  # best performance per dollar

Claude Code Skill Skill

# In Claude Code, use the RightRank skill:
/rightrank best model for terminal-bench-hard
/rightrank compare top 5 by value
/rightrank cheapest model scoring above 80 on terminal-bench-hard