LEASH Router Matches Claude Opus 5 on OSWorld at 19.2% Less Cost

A new training-free routing framework called LEASH gets a 35B-parameter Qwen model to 73.66% task success on OSWorld at $0.4117 per task, close enough to Claude Opus 5's 72.90% to call it a tie, while cutting inference cost by 19.2%, according to a paper posted to Hugging Face.

The setup is to let a weak open-weight agent act by default and hand control to a frontier model only when the weak agent's own hidden states say it is drifting toward failure. 'We compare the weak model's hidden states with those from its past successful and failed trajectories, and a state that looks more like the failures signals higher uncertainty,' the authors write. Because the estimator reuses activations already produced during action generation, it needs no extra forward pass or dedicated monitor training.

Even after handoff, a 'shadow observer' stays on the strong agent, which matters because the strong agent is usually served from a black-box API that exposes no internals. Across four model backbones on OSWorld, the authors say LEASH 'outperforms random escalation by 6.55--11.94 points at matched escalation rates.'

The abstract names only Qwen3.5-35B-A3B among the four backbones, and does not break out per-task-category results or shadow-observer catch rates. Agent work has been dense in our recent tracker; this is one of several October papers pointing at cost-aware orchestration rather than another frontier model.