Pick the model.
Keep the browser.
One agent harness. A choice of models. Find the right balance for the work you actually run.
Managed agent models
Per 1M tokens · USD
| Model | Input | Output |
|---|---|---|
| Claude Opus 5 | $6.00 | $30.00 |
| Claude Sonnet 5 | $2.40 | $12.00 |
| Grok 4.5 | $2.40 | $7.20 |
| GPT-6 Astra | $12.00 | $60.00 |
| GPT-5.6 | $6.00 | $36.00 |
| GPT-5.6 LunaDEFAULT | $0.24 | $1.44 |
| Gemini 3.6 Flash | $1.80 | $9.00 |
| MiniMax M3 | $0.36 | $1.44 |
Choose per run
Start with GPT-5.6 Luna, the default. Change the model as your task's accuracy, latency, and cost requirements change.
Model documentationUse your own key
Eligible plans support your model-provider key. Provider token costs, Browser Use orchestration, browser time, and traffic remain separate charges.
Understand the chargesPurpose-built models
Browser Use also develops models trained for browser work. BU 2.0 has its own model API rates, separate from managed V4 runs.
Purpose-built model pricingChoose with evidence
Compare on browser work.
Accuracy and cost depend on the task, model, and harness. Start with a matched benchmark, then test your own workflow.
Browser Use: 82% of tasks solved at 17¢ each. That is 20 points better than Opus 5, which costs 20× more per solved task.
Earlier Browser Use 1.0 latency measurements use the OnlineMind2Web workload. Read that experiment for its model versions and methodology.