Browser Agent hits 88% success on BU Bench and beats Browser Code

Cutting token usage by 91% is the only way we're actually going to make browser agents viable for daily use. Most current agentic workflows are just too expensive and slow because they bloat the context window with unnecessary DOM data, which kills the VRAM budget and drives up costs. I've been looking into a browser agent harness that fundamentally changes this by prioritizing token efficiency, and the results on the BU Bench v1 benchmark are pretty staggering compared to Browser Code.

The performance gap isn't just a marginal gain; it's a significant jump in reliability and speed. Here is how the two stack up:

• Success Rate: Browser Agent hit 88% while Browser Code trailed at 78%.
• Cost: Browser Agent cost $5.37 compared to $8.34 for Browser Code.
• Execution Time: Browser Agent clocked in at 32,694 seconds, whereas Browser Code took 47,970 seconds.

The real technical win here is the reduction in compute requirements. By stripping away the noise and focusing on token efficiency, it becomes feasible to fine-tune a small, specialized model rather than relying on a massive frontier LLM for every single click or keystroke. Since browser agent inference is intermittent—meaning the model thinks, acts, and then waits for the page to load—a small, distilled model can handle each step with minimal GPU time.

This shifts the entire economic model of AI agents. Instead of the standard token-based pricing we see from the big labs, which makes costs unpredictable and often prohibitively high for complex tasks, a specialized model allows for per-task pricing. This is a huge advantage for anyone trying to build a scalable AI workflow because it makes the overhead predictable.

For those interested in the architecture or the specific benchmarks, the full breakdown is detailed here:

https://www.pierrebarreau.com/blog/improving-the-state-of-the-art-in-agentic-browsing

If you're into prompt engineering or building LLM agents, the move toward specialized, small-footprint models for specific environments like the browser is definitely the right direction. We don't need a trillion-parameter model to tell a browser to click a "Submit" button; we need a lean, fast model that understands the DOM without eating $10 in tokens per session. This approach proves that optimization in the harness can actually outperform raw model power.