Google’s BATS lifts BrowseComp accuracy to 24.6% – cuts agent cost 31%

Google’s new BATS framework is the first agent paper this week that feels directly aimed at your cloud bill. The lightweight Budget Tracker plugin alone hits ReAct‑level accuracy on web tasks using 10 instead of 100 tool calls and trims overall cost by about 31%, simply by exposing live “query budget remaining” counters inside the agent’s reasoning loop.

Full BATS orchestration pushes harder on quality. On BrowseComp, a Gemini‑2.5‑Pro agent with BATS scores 24.6% vs ReAct’s 12.6% under the same 100‑tool cap; BrowseComp‑ZH jumps to 46.0% vs 31.5%, and HLE‑Search to 27.0% vs 20.5%. Planning and self‑verification both become budget‑aware, and the paper introduces a unified cost metric that blends token spend and tool‑call prices so you can reason in dollars, not abstract “steps.”

DAIR.AI is already teaching BATS in its agent courses, which is usually a sign a pattern is graduating from research toy to production norm. Paired with Artificial Analysis’ token‑usage charts showing GPT‑5.2‑xhigh burning nearly 2× Sonnet’s tokens for similar work, the direction of travel is clear: agents that ignore budgets are going to feel as dated as models that ignore context windows. Time to make “budget‑aware by design” part of your standard agent spec.