Fastbrowse bets on a “fast” model for browsing: 11 times cheaper, but the benchmark is the company's own

By Hacker News (Show HN) · October 30, 2026.

Furqan, founder of Nebula, has presented Fastbrowse on Hacker News, an open-source browser agent that controls Chrome from a natural-language task. The code is on GitHub (agent-labs-dev/fastbrowse) and, according to its author, it launches with a single line using uv, Chrome and a TypeSafe or OpenRouter key.

The design rests on two key decisions. First, at each step the agent receives the list of elements on the page instead of screenshots. According to the author, this stops it from trying to click buttons that don't exist. Second, it uses what Nebula calls "System One models" (currently one called Jev), designed for fast, cheap decisions, by analogy with Kahneman's fast, intuitive thinking. If Jev fails, the system falls back to a low-cost LLM such as Gemini Flash Lite or DeepSeek v4.1.

The figures, according to Nebula itself: on a set of 46 tasks run three times each against Browser Use's hosted agent with its default configuration, both on Browser Use cloud browsers, Fastbrowse cost $0.030 per run versus $0.318 (the "11 times cheaper" they advertise). It passed 137 of 138 runs; its rival, 133. We don't have access to the task details or to an independent replication.

At the time of publication the post had only 3 points and 2 comments, so there has been no community scrutiny yet.

**Our take.** The idea is sensible and fits a clear trend: browser agents are expensive and slow because they use a large model and vision on every click, and many of those steps are routine. Routing the easy steps to a small model and escalating to a larger one only when it fails is the logical architecture for making web automation affordable. Reading the page structure instead of pixels is also more robust, though it has its limits: pages with canvas, complex dynamic content or poorly labeled elements can hide what the agent needs.

Enthusiasm should be tempered. A 46-task suite designed by the tool's own authors, compared against a competitor in its default configuration, is an indication, not a measurement. The difference in successes (137 vs. 133 of 138) is small and could be down to task selection. The most reliable figure is the cost, because it's the hardest to dress up, but it's worth testing with your own workflows.

If confirmed, it matters for more than the savings: when the cost per task drops by an order of magnitude, automating paperwork, research and repetitive errands goes from a luxury to an everyday thing. In the short term, that puts pressure on repetitive administrative work, the most exposed according to our coverage of jobs. In the long term, it's the kind of cheap, open infrastructure that democratizes automation. And let's remember the perennial question with agents: who controls the credentials and permissions of a bot that browses on your behalf.