Which LLM Wins the Agent Benchmark? PinchBench Success, Speed, and Cost Rankings Revealed

OpenClaw founder Peter Steinberger introduced PinchBench, a benchmark that ranks 32 mainstream large language models on three dimensions—success rate, execution speed, and cost—specifically for AI agent workloads.

Choosing a model is the highest‑leverage decision for an agent system.

Success Rate Rankings

google/gemini-3-flash-preview – 95.1% success, ranking first.

minimax-m2.1 – 93.6% success, surpassing Claude Sonnet 4.5 (92.7%) and GPT‑4o (85.2%).

moonshotai/kimi-k2.5 – 93.4% success, third place; two of the top three are Chinese models.

The benchmark measures the percentage of tasks successfully completed in a standardized OpenClaw agent test. Tasks cover practical work such as scheduling meetings, prioritizing emails, writing code, and managing files, rather than trivia or math problems.

Speed Rankings

minimax/minimax-m2.5

google/gemini-2.0-flash

meta-llama/llama-3.1-70b

Cost Rankings

openai/gpt-5-nano

google/gemini-2.5-flash-lite

mistralai/devstral-2512

Overall Assessment

MiniMax‑M2.1 and Kimi‑K2.5 provide the best cost‑performance balance, while the Claude Opus series is comparatively expensive.

GPT‑5‑Nano: Cost #1, Success #9, Speed #16

Gemini‑2.5‑Flash‑Lite: Cost #2, Success #14, Speed #13

MiniMax‑M2.1: Cost #5, Success #2, Speed #22

Kimi‑K2.5: Cost #8, Success #3, Speed #27

Claude Opus 4.6: Cost #20, Success #7, Speed #30

Traditional "intelligence" benchmarks such as MMLU and HumanEval increasingly fail to predict which models can effectively operate as agents; agent tasks demand abilities like multi‑step instruction following, tool invocation, handling ambiguous scenarios, and error recovery, which differ fundamentally from simple question‑answering.