Summary: On October 31, 2026, Browser Use opened BU Ultrafast to every Browser Use Cloud plan in the US, ending its waitlist. The median completed task takes 22 seconds and costs $0.0064, versus 79 seconds and $0.0098 on their default agent. For anyone shipping browser automation, the migration cost is a single model string.
What actually shipped
Until October 30, BU Ultrafast sat behind a waitlist. According to the announcement by co-founder and CTO Gregor Zunic, it is now the fastest agent on all US Cloud plans, including the free tier. If you were waitlisted, you already have it — no action needed.
The numbers come from completed-run medians between October 16–29, 2026, capped at 1,000 runs per project: 12,238 Ultrafast runs across 134 projects, against 146,120 default runs across 9,473 projects. In the 8 projects with 20+ runs on both presets, Ultrafast was faster and cheaper in all 8. One caveat worth reading: cost covers model and search charges only — browser and proxy time are billed separately, so your real per-task economics will differ.
The change is one line
You switch by setting model to bu-ultrafast:
curl https://api.browser-use.com/api/v4/runs \
-H "X-Browser-Use-API-Key: $BROWSER_USE_API_KEY" \
-H "Content-Type: application/json" \
-d '{"task": "Find the cheapest flight from SFO to JFK next Friday", "model": "bu-ultrafast"}'Nothing else in your run config changes. If a task needs more reasoning per step, the same docs point you to bu-fast — same agent, medium reasoning, standard tier. For research, bulk collection, or long instructions, omit the model entirely and get the GPT-6 Luna default with High reasoning.
Tradeoffs before you flip everything
Speed is not free. Per the docs, bu-ultrafast runs GPT-6 Luna on OpenAI’s priority tier, billed at 2x the standard rate — so the per-task median is cheaper, but your token economics shift. bu-fast bills at standard rates, as does bu-ultrafast on HIPAA projects.
The presets are also strict. They require Browser Use Pi, accept no modelParams, mode, or harness: "browsercode", and reject any such field with a 422. The preset sticks to the session: a follow-up that sets a different model or model params also fails. And new sessions run without the managed proxy unless you set browserSettings.proxyCountryCode yourself.
This fits a pattern I keep noticing in agent tooling: instead of one agent tuned by config knobs, vendors ship opinionated presets with hard edges. I looked at the same idea from the other side in Browserbase’s managed browser agents — fewer degrees of freedom, more predictable runs. Here it’s explicit: if you send a forbidden field, the API refuses rather than silently blending behavior.
A practical next step
If you have short, interaction-heavy flows — form fills, lookups, quick extractions — they’re the natural first candidates. Pull your slowest task cohort, rerun it on bu-ultrafast, and compare completed-run medians on your own workload rather than the vendor’s. Watch the excluded costs (proxy and browser time) and the 422s from preset restrictions. If tasks stall on reasoning-heavy steps, bu-fast is the middle setting without changing anything else.
Sources
View the original post on X