When we started, GPT-4o was state of the art for browser agents. Since then, we improved our library a ton and released BU 1.0. Below is a comparison of GPT-4o and BU 1.0 with latest version of our library.
We track three core metrics for our browser agents: accuracy, speed, and cost. In one year, the accuracy improved from 71.8% to 82.0%. But more importantly, we reduced average task time from 123 seconds to 33.4 seconds and dropped the cost from staggering 39.2¢ to 1.9¢ per task. It now costs less than 2 cents to complete a simple browsing task.
What’s Next?
We’re already at <2¢ per task and 4x faster than a year ago. The cost and speed problems are largely solv…
When we started, GPT-4o was state of the art for browser agents. Since then, we improved our library a ton and released BU 1.0. Below is a comparison of GPT-4o and BU 1.0 with latest version of our library.
We track three core metrics for our browser agents: accuracy, speed, and cost. In one year, the accuracy improved from 71.8% to 82.0%. But more importantly, we reduced average task time from 123 seconds to 33.4 seconds and dropped the cost from staggering 39.2¢ to 1.9¢ per task. It now costs less than 2 cents to complete a simple browsing task.
What’s Next?
We’re already at <2¢ per task and 4x faster than a year ago. The cost and speed problems are largely solved. But there’s one critical dimension where browser agents still fall short: reliability.
Reliability is still the biggest pain point preventing widespread adoption. 2025 has been the year of agents, but 2026 needs to be the year of reliable agents—agents that either complete tasks successfully or fail transparently with clear diagnostics. We are working extremely hard towards not only increasing the accuracy but also the confidence calibration of our models to enable better observability.
We believe no human should waste time on repetitive work and our goal is to make browser automation ubiquitous in our second year.