Jake Benchmark v1 crowns Qwen3.5 27B

The real takeaway is that local agent quality still depends more on tool discovery and execution discipline than on raw parameter count. A quantized 27B beating a 35B by 2.5x is a loud reminder that "bigger" is often the wrong instinct for agent workloads.

• –Medium thinking was the sweet spot for the winner; higher reasoning hurt, which suggests agent loops need crisp tool selection, not endless deliberation.
• –Models that found the `gog` CLI did the work; models that couldn't discover it mostly died under 5%, making tool surfacing the biggest bottleneck.
• –Browser automation was a complete bust across the board, so end-to-end UI agents still need more than a good language model.
• –Security behavior was all over the place, from clean phishing refusal to reckless secret probing, so guardrails can't rely on model instinct alone.
• –OpenClaw advertises Gmail, GitHub, and browser integrations ([openclaw.ai](https://openclaw.ai/)), but this benchmark suggests the model remains the limiting layer.