devdot
← All postsEngineering ·

Browser Use Made a 15x Pricier Model the Default. Pin Your Model Names in Code.

Browser Use 1.0.6 switched its default agent model to bu-4.0, which costs 15 times as much per token as bu-fast. Here is how to stop SDK defaults from quietly making pricing decisions for your team.

Browser Use shipped version 1.0.6 on November 16, and one line in the release notes matters more than the rest. Agent(task) and ChatBrowserUse() now run bu-4.0 by default instead of bu-fast.

The new model costs $1.80 per million input tokens and $9.00 per million output. The old default cost $0.12 and $0.60. That's 15 times the price, and the release notes are upfront about it: a test run that used to cost $0.0014 now costs $0.022.

To be fair to the Browser Use team, they announced it clearly and gave a one-line escape hatch. Pass llm='bu-fast' and you're back where you were. The problem is everyone who never reads release notes, which is most teams running a scheduled job with a loose version range in their requirements file.

How a default turns into an invoice

Here's the usual path. Someone prototypes a browser agent, it works, and the code ends up looking something like this:

agent = Agent(task="Check competitor pricing pages")

No model named. The library picks. Months later a dependency bot bumps the package, CI passes because the agent still completes the task (probably better than before), and the change merges on a Friday. Nothing fails. The only signal is the bill at the end of the month, or a spend alert if you were disciplined enough to set one.

Tests won't catch this, because nothing is broken. A more capable model passes your tests more often. Quality going up is exactly what makes the cost change invisible.

Treat the model name as config you own

The fix is boring and takes five minutes.

Name the model everywhere. Every agent, every client constructor, every call. If the SDK has a default, act as though it doesn't exist. Put the model name in config so changing it is a reviewed diff, not a side effect of pip install -U.

Pin the library version exactly for anything that runs unattended. browser-use==1.0.5, not >=1.0. Upgrade on purpose, read the changelog, then bump.

Add a cost assertion to one smoke test. Run a known task, read the token usage from the response, and fail the build if cost per run jumps by more than, say, 2x. It's a crude check and it would have caught this one in seconds.

assert run.cost_usd < BASELINE_COST * 2, f"Cost jumped to {run.cost_usd}"

Alert on spend per job, not per account. A monthly account cap tells you something went wrong somewhere. A per-workflow budget tells you which workflow, on the day it happened.

This will keep happening

Browser Use isn't unusual here. Vendors have every reason to move defaults toward their newest, most capable model. Newer models make the product look better in demos and benchmarks, and a default is the strongest lever they have. Expect more SDKs, gateways and agent frameworks to do the same over the next year, and not all of them will put the price in the release notes.

We'd happily pay for bu-4.0 where it earns its keep. Some browser tasks really do need the stronger model, like long multi-step forms or pages that fight automation. Plenty don't. Scraping a pricing table every morning probably doesn't need 15x the spend. That should be a decision someone on your team makes with the numbers in front of them, and a library update shouldn't make it for you.

So go grep your codebase for agent and client constructors with no model argument. If you find any, that's your next ticket.

We're here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you're looking to build something, get in contact with us today!

NEXT POST →AI Assistants Sent Shoppers to Fake Stores in Up to 15% of Answers. Tell Them Where Your Real One Is.