TAU-bench

Agentic

Tool-Augmented Understanding benchmark

Source ↗ Higher is better

Rank	Model	Provider	Score
1	Qwen: Qwen3.5 397B A17B	Alibaba	87.9%
2	Gemini 3.0 Pro	Google	85.4%
3	Anthropic: Claude Opus 4.5	Anthropic	85.3%
4	OpenAI: GPT-5.2	OpenAI	84.8%
5	Anthropic: Claude Sonnet 4.5	Anthropic	84.7%
6	Gemini 3 Flash	Google	83.5%
7	gemini-3-pro	Google	82.5%
8	Qwen: Qwen3 Max ThinkingHistorical	Alibaba	82.2%
9	Z.ai: GLM 5	Z.ai	81.0%
10	DeepSeek: DeepSeek V3.2Historical	DeepSeek	80.4%
11	OpenAI: GPT-5	OpenAI	80.0%
12	Nemotron-Orchestrator-8B	NVIDIA	76.3%
13	Qwen: Qwen3 MaxHistorical	Alibaba	72.0%
14	MoonshotAI: Kimi K2 0711	Moonshot AI	64.3%
15	Claude-3.7-Sonnet	Anthropic	61.8%
16	OpenAI: o4 Mini	OpenAI	56.9%
17	OpenAI: GPT-4.1	OpenAI	54.7%
18	OpenAI: GPT-4.1 Mini	OpenAI	53.0%

Access this data

JSON GET /api/v1

[
  {
    "rank": 1,
    "model": "Qwen: Qwen3.5 397B A17B",
    "provider": "Alibaba",
    "score": "87.91333333333334"
  },
  {
    "rank": 2,
    "model": "Gemini 3.0 Pro",
    "provider": "Google",
    "score": "85.43333333333334"
  },
  {
    "rank": 3,
    "model": "Anthropic: Claude Opus 4.5",
    "provider": "Anthropic",
    "score": "85.31"
  }
]

Curl

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

CLI brew install rightrank

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

Claude Code Skill Skill

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