Reasoning model uses tools on only 24% of reachable tasks

A reasoning model called an available tool on only 23.9% of tasks where a tool could help, per a new OSWorld-MCP benchmark. Across the same 309 tasks, the toolset lifted the reasoning model's score by 4 points and dropped a non-reasoning model's score by 5.9 points. Wiring tools into an agent isn't automatically a win. It depends which model reads the tool list.

Training changes the numbers more than the model does. A dense tool-use bonus added during multi-turn reinforcement learning pushed spreadsheet tool adoption from 3% to 33%, the paper's authors found. Held-out task accuracy didn't move. The model got better at reaching for the tool, not at solving the task. Those are two separate claims that keep getting reported as one.

The cheaper win

The paper's other result has nothing to do with model behavior. Dropping the screenshot right after a successful tool call, and halving how much image history the agent carries, cut input tokens by about a third. Retrained under that rule, the trimmed agent scored 37.8% against 33.0% for the uncompressed version, at 53% of the input token cost. That's a context-management fix, not a smarter model, and it's the only change in the paper that actually moved accuracy.

I've hit the same wall building my own agents. Stuffing every screenshot and tool result into context feels safe until the token bill shows up. This paper backs up trimming stale state aggressively before reaching for a bigger model.

Tool-adoption rate and task accuracy move independently here. Any agent benchmark reporting one without the other is telling you less than it sounds like.