Show HN: Browser Agent – cost efficient browser automation

A developer's Browser Agent harness outperforms Browser Code on the BU Bench v1 benchmark, achieving an 88% success rate at $5.37 and 32,694 seconds, compared to Browser Code's 78% success rate at $8.34 and 47,970 seconds. The harness uses 91% fewer tokens, aiming to reduce compute and VRAM costs for fine-tuning a small specialized model, and enables per-task pricing as a counter-positioning strategy against token-based pricing from frontier labs.

I've built a browser agent harness that outperforms Browser Code on their own benchmark BU Bench v1 , on success rate, speed and cost. Browser Agent: 88% success rate, $5.37, 32,694 seconds Browser Code: 78% success rate, $8.34, 47,970 seconds For browser agents to become ubiquitous, they need to be faster, cheaper, and more reliable. The harness was built with token efficiency in mind and uses 91% fewer tokens than Browser Code to reduce the compute and VRAM budget necessary to fine tune a small, specialised model. Because browser agent inference is intermittent, a small specialised model could serve each step with minimal GPU time. This makes per-task pricing viable, offering users cheaper, more predictable costs and giving the operator a counter-positioning advantage over frontier labs pricing their models by token usage. More about the comparison between Browser Agent and Browser Code here: https://www.pierrebarreau.com/blog/improving-the-state-of-th... https://www.pierrebarreau.com/blog/improving-the-state-of-the-art-in-agentic-browsing If anyone is interested in fine tuning that model, would love to talk Comments URL: https://news.ycombinator.com/item?id=49257523 https://news.ycombinator.com/item?id=49257523 Points: 3 Comments: 0