Skip to main content

PinchBench

An OpenClaw agent benchmark from Kilo that measures successful task completion across standardized real-world agent workflows.

How BenchLM shows PinchBench

BenchLM mirrors the public PinchBench average-success-rate view using the official snapshot updated on null: null models and 559 runs. PinchBench grades runs with automated checks plus an LLM judge.

This benchmark is display only on BenchLM. It is excluded from BenchLM overall rankings, category rankings, and weighted scoring. The table below uses average scores only, matching the public PinchBench average view rather than the best-run view.

null models559 runsAverage scores onlyOfficial runsDisplay only

Average success rate on PinchBench — June 2, 2026

BenchLM mirrors the published average success rate view for PinchBench. Qwen3.7 Max leads the public snapshot at 92.5% , followed by nvidia/nemotron-3-ultra-550b-a55b (89.9%) and Grok Build 0.1 (88.9%). BenchLM does not use these results to rank models overall.

41 modelsAgenticCurrentDisplay onlyUpdated June 2, 2026

The published PinchBench snapshot is tightly clustered at the top: Qwen3.7 Max sits at 92.5%, while the third row is only 3.6 points behind. The broader top-10 spread is 13.3 points, so the benchmark still separates strong models even when the leaders cluster.

41 models have been evaluated on PinchBench. The benchmark falls in the Agentic category. This category carries a 22% weight in BenchLM.ai's overall scoring system. PinchBench is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About PinchBench

Year

2026

Tasks

23 OpenClaw agent tasks

Format

Average success rate from official runs

Difficulty

Long-horizon agent workflows

PinchBench publishes official OpenClaw runs across 23 tasks and grades results with automated checks plus an LLM judge. BenchLM mirrors the public average-score view as a display-only benchmark.

BenchLM freshness & provenance

Version

PinchBench 2026

Refresh cadence

Quarterly

Staleness state

Current

Question availability

Public benchmark set

CurrentDisplay only

BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.

Average success rate table (41 models)

1
Qwen3.7 Maxqwen/qwen3.7-max
92.5%
3
Grok Build 0.1x-ai/grok-build-0.1
88.9%
4
82.6%
5
DeepSeek V4 Flashdeepseek/deepseek-v4-flash
81.7%
6
Gemini 3.1 Progoogle/gemini-3.1-pro-preview
81.0%
7
Gemini 3.1 Flash-Litegoogle/gemini-3.1-flash-lite
80.5%
8
Grok 4.20x-ai/grok-4.20
80.3%
9
Step 3.5 Flashstepfun/step-3.5-flash
79.4%
10
GPT-5.4 miniopenai/gpt-5.4-mini
79.2%
11
Claude Opus 4.7anthropic/claude-opus-4.7
76.0%
12
GPT-5.4openai/gpt-5.4
75.7%
13
Seed-2.0-Litebytedance-seed/seed-2.0-lite
75.0%
14
Gemini 3.5 Flashgoogle/gemini-3.5-flash
74.2%
15
Grok 4.3x-ai/grok-4.3
73.7%
16
Claude Opus 4.6anthropic/claude-opus-4.6
72.5%
17
Qwen3.6 Plusqwen/qwen3.6-plus
72.5%
18
Gemini 3 Flashgoogle/gemini-3-flash-preview
72.1%
19
GLM-5-Turboz-ai/glm-5-turbo
71.8%
21
Claude Haiku 4.5anthropic/claude-haiku-4.5
67.7%
22
GLM-5V-Turboz-ai/glm-5v-turbo
67.6%
23
MiniMax M2.7minimax/minimax-m2.7
66.8%
24
Trinity-Large-Thinkingarcee-ai/trinity-large-thinking
65.7%
26
Claude Sonnet 4.6anthropic/claude-sonnet-4.6
62.7%
27
DeepSeek V4 Prodeepseek/deepseek-v4-pro
61.1%
29
Kimi K2.5moonshotai/kimi-k2.5
54.6%
31
Trinity-Large-Thinkingarcee-ai/trinity-large-preview
53.0%
34
GPT-OSS 120Bopenai/gpt-oss-120b
44.8%
35
Nemotron 3 Super 120B A12Bnvidia/nemotron-3-super-120b-a12b
42.2%
37
GPT-OSS 20Bopenai/gpt-oss-20b
36.3%
38
Mercury 2inception/mercury-2
32.7%
39
GPT-5.5 Proopenai/gpt-5.5-pro
21.4%
40
Llama 3.1 70B Instructmeta-llama/llama-3.1-70b-instruct
10.7%
41
Llama 4 Scoutmeta-llama/llama-4-scout
3.2%

FAQ

What does PinchBench measure?

An OpenClaw agent benchmark from Kilo that measures successful task completion across standardized real-world agent workflows.

Which model leads the published PinchBench snapshot?

Qwen3.7 Max currently leads the published PinchBench snapshot with 92.5% average success rate. BenchLM shows this benchmark for display only and does not use it in overall rankings.

How many models are evaluated on PinchBench?

41 AI models are included in BenchLM's mirrored PinchBench snapshot, based on the public leaderboard captured on null.

Last updated: June 2, 2026 · mirrored from the public benchmark leaderboard

The AI models change fast. We track them for you.

For engineers, researchers, and the plain curious — a weekly brief on new models, ranking shifts, and pricing changes.

Free. No spam. Unsubscribe anytime.