DIRTY STATIC / DISPATCH Tools16 NOV 2026 • 9 MIN READ
A robotic arm at a keyboard facing a large glowing caution yellow browser window on a concrete wall in a brutalist hall, hazard stripes on the floor
ToolsINTELLIGENCE FOR THE MACHINE AGE

Browser Use 1.0.6: BU 4.0 Becomes Default

Browser Use 1.0.6, checked on November 16, 2026, makes BU 4.0 the default model for the Python browser agent library when no model is named, and moves the default reasoning setting from medium to high. This profile covers what the change does to code that omits the model, the full BU 4.0 price ladder beside BU Fast, the separate browser and traffic charges, the new engine under the 1.0 line, and how to keep the cheaper model on purpose.

Start reading ↓

A quiet upgrade can change a bill without changing a line of your code. That is the story in Browser Use 1.0.6. Code that never named a model used to run on BU Fast through the Browser Use gateway. From 1.0.6, that same code runs on BU 4.0, and it thinks harder by default while it does it.

Browser Use is an open source library that lets an AI model drive a real browser. You give the agent a task in plain words, and the model opens pages, clicks, types and reads results until the task is done or the run gives up. Version 1.0 rebuilt the engine only days ago. Version 1.0.6 changes which brain the agent uses when you do not choose one yourself. This article is based on the official release notes, the official migration guide and the official model and pricing pages, checked on November 16, 2026. No run by this desk is claimed.

Browser Use

Official site

An open source browser agent library. Primary use case: giving a language model a real browser so it can complete web tasks from a written instruction, in your own Python code, through the CLI, or through the hosted cloud.

Key facts in 1.0.6: the default model is now bu-4.0 on the Browser Use gateway when no model is named; the default reasoning setting is high, up from medium; bu-4.0-mini and bu-4.0-pro sit below and above it in the same family; and your own model from another provider still works when you pass it in.

Advantages: one Browser Use key runs the model, the cloud browsers, web search, the CLI and the MCP server, and the library itself costs nothing to use. Limitations in this change: a default is now a spending decision, BU 4.0 and BU 4.0 Pro are listed as US only for now, and the total bill adds browser time and traffic on top of token charges.

Platforms
Python library; Browser Use Cloud browsers, or a local Chromium with Browser('chromium')
Pricing
Library free under MIT; gateway models, cloud browser time and traffic billed by usage

What changed, in one move

The change is narrow and total at the same time. It is narrow because only the default moved. Code that names its model sees no change from this release on that point. It is total because a large share of real code never names a model. Tutorials omit it. Prototypes omit it. Plenty of production jobs omit it too, because the default worked and nobody had a reason to touch it.

From 1.0.6, an omitted model means bu-4.0, served through the Browser Use gateway, with reasoning set to high unless you say otherwise. The migration guide states the same fact in its comparison table: the model row for 1.0 now reads bu-4.0, with a note that it was bu-fast before 1.0.6. The old gateway models from the 0.x line, such as bu-2-0 and bu-1-0, do not carry over. They raise an error that points you at bu-4.0 or back at the 0.x line.

The reasoning move matters as much as the model move. High reasoning means the model spends more tokens thinking through each step. On a hard task, that can be the difference between a run that recovers from a strange page and a run that loops. On a simple task, it can be money spent thinking about a form with three fields. The release gives you the stronger setting by default and leaves the cheaper setting one argument away.

Python

agent = Agent(task, llm='bu-fast')

That one line is the whole opt out. Name bu-fast in code and the run keeps the old default model. Name nothing and you have accepted the new one, whether you read the release notes or not.

The price ladder, checked today

Token prices are where this change lands. The official model page, checked on November 16, 2026, lists the BU 4.0 family like this, in US dollars per million tokens.

ModelAPI stringInputCache readOutput
BU 4.0 Minibu-4.0-mini$0.20$0.02$1.00
BU 4.0bu-4.0$1.80$0.09$9.00
BU 4.0 Probu-4.0-pro$3.00$0.15$15.00
BU Fastbu-fast$0.12$0.012$0.60

Read the last two rows together and the shape of 1.0.6 is clear. BU 4.0 costs 15 times BU Fast on both input and output at the listed rates. That is a rate ratio, not a bill forecast. A task that finishes in fewer steps, with less retry and less wandering, can close part of that gap. A task that thinks at high reasoning through many steps can widen it. Anyone who quotes you a clean multiple for a finished task is guessing.

The project offered one small example of its own. In the release coverage, Browser Use reported one short test at $0.022 on bu-4.0 and $0.0014 on bu-fast. Treat that as one vendor example on one short task, not as a rate card for your workload. Your tasks have their own lengths, their own failure modes and their own retry counts.

Tokens are also not the whole bill. The model page states that browser sessions and network traffic are charged apart from tokens. It lists browser sessions at $0.06 per hour, with traffic at $5 per GB through the managed proxy or $0.20 per GB without it or with your own proxy. A long run in a cloud browser pays on all three meters at once.

A large industrial scale in a concrete room weighing a small browser window against a much larger one, hazard stripes and a caution yellow lamp Two browser windows on a scale, one far heavier than the other. The listed token rates for the new default run 15 times the old one. Illustration: Dirty Static

The engine this default sits on

Version 1.0.6 is a small release on top of a large rewrite, and the rewrite explains why the default matters more now. In the 0.x line, the model read a dump of the page and picked one action per step, such as a click on an indexed element. In 1.0, the model writes Python in a persistent REPL, a live coding session that keeps its state between steps. That session runs with helpers from the project and with raw CDP, the Chrome DevTools Protocol, which is the interface Chrome itself exposes for control and inspection.

One written cell can click, type, read and loop in a single step, without waiting for a built in action to exist for each move. The agent loop runs on Pydantic AI, and the common surface stayed familiar through the rewrite. Agent, Browser, Tools, run() and final_result() still work, which is why so much existing code upgrades with little friction and then meets the new default without noticing.

The browser default changed in the same rewrite, and it belongs in any cost talk about 1.0.6. A plain Browser() now starts a Browser Use Cloud browser, with stealth features, proxies and CAPTCHA solving handled by the service, and nothing to install on your machine. A local browser is still there when you ask for it by name, as Browser('chromium'). Tasks that point at localhost or a private address need that local route, because a cloud browser cannot reach them. The migration guide says so in plain terms and warns once when it sees such a task.

For readers weighing this tool against a direct automation library, our Playwright 1.65 profile covers the other end of the trade, where your code names every step and no model bill arrives at all.

What Works

  • The stronger model is now the path of least resistance for hard web tasks
  • The opt out is one explicit argument, and naming a model was always the safer habit
  • One key covers the model, cloud browsers, search, the CLI and the MCP server
  • The library stays free under MIT, and own provider models still work

What Does Not

  • Listed token rates for the new default run 15 times the old default
  • High reasoning by default spends more tokens on simple tasks that never needed it
  • A silent upgrade can raise a production bill before anyone reads the notes
  • BU 4.0 and BU 4.0 Pro are listed as US only for now, and browser time and traffic bill apart from tokens

Which should you choose

Keep BU 4.0 if your tasks are hard and failure is the real cost. A research run across messy sites, a flow with logins and changing layouts, or a task where a failed run wastes human time can justify the stronger model at high reasoning. Judge it per completed task, not per token.

Pin BU Fast if your tasks are simple and repeat often. Form fills, checks on pages you know well, and high volume runs with a clear success test are the natural home of the cheaper model. The saving is real at the listed rates, and the model was good enough to be the default until this week.

Try BU 4.0 Mini in the middle. At $0.20 input and $1.00 output per million tokens, Mini sits far closer to BU Fast on price while staying inside the new family. For many routine tasks it may be the sensible default that this release did not pick.

Name your model either way. That is the durable lesson of 1.0.6. A default is a decision the vendor can revisit in any release, and this vendor revisits often. Code that states its model, its reasoning level and its browser choice keeps its behaviour and its bill shape when the next point release lands.

Final thoughts

Browser Use 1.0.6 does not add a feature you can point at. It changes what happens when you point at nothing. The new default is stronger, slower to be cheap, and defensible for the hard tasks this tool is bought for. It is also a reminder that in agent software, the most important line in a release is often the one about what changed while you were not looking. Pin the version, name the model, and measure cost per finished task on your own work before you let any default decide for you.

Sources

■

END OF TRANSMISSION

Stay curious. Question everything.

CONTINUE THE DESCENT

MORE FROM DIRTY STATIC ↘