BROWSER USE

Products:
- [Web Agents](https://browser-use.com/web-agents)
- [Stealth Browsers](https://browser-use.com/stealth-browsers)
- [Custom Models](https://browser-use.com/custom-models)
- [Proxies](https://browser-use.com/proxies)
- [Skill APIs](https://browser-use.com/skills)

[Pricing](https://browser-use.com/pricing)
[Blog](https://browser-use.com/posts)
[Cloud Docs](https://docs.cloud.browser-use.com)
[Open Source Docs](https://docs.browser-use.com)

[GET STARTED](https://cloud.browser-use.com)
[GITHUB](https://github.com/browser-use/browser-use)

---

# What LLM model should I use for Browser Use? The Definitive Browser AI Benchmark

**Author:** Alexander Yue
**Date:** 2026-02-19
> A statistically rigorous benchmark comparing the speed, cost, and accuracy of top frontier models for browser automation.

---

![BU Bench V1: Success Rate by Model](https://browser-use.com/images/benchmark/accuracy_by_model_dark.png)

One question dominates every conversation we have with developers: *"Which model should I plug into Browser Use?"*

A single test run means nothing on the chaotic, ever-changing real web. We run our benchmark multiple times per model and aggregate with statistical bootstrapping for real error bars. Our [evaluation system](https://browser-use.com/posts/sota-technical-report) has run over 600,000 tasks, validated by an LLM judge achieving 87% agreement with human labels.

## The results


| Model | Type | Score |
| --- | --- | --- |
| Browser Use Cloud (bu-ultra) | Cloud | 78.0% |
| OSS + BU LLM (ChatBrowserUse-2) | OSS + Cloud LLM | 63.3% |
| claude-opus-4-6 | Open Source | 62.0% |
| gemini-3-1-pro | Open Source | 59.3% |
| claude-sonnet-4-6 | Open Source | 59.0% |
| gpt-5 | Open Source | 52.4% |
| gpt-5-mini | Open Source | 37.0% |
| gemini-2.5-flash | Open Source | 35.2% |


"Open Source" means running the open-source Browser Use library with that LLM. "OSS + Cloud LLM" means the open-source library with our ChatBrowserUse-2 model. "Cloud" is the fully managed Browser Use Cloud agent. You can view all models and their prices [here](https://docs.browser-use.com/supported-models#available-models).

100 hand-selected tasks from five sources (Custom, WebBench, Mind2Web 2, GAIA, BrowseComp). Each task is hard but verified completable. Full methodology in our [benchmark post](https://browser-use.com/posts/ai-browser-agent-benchmark).

![BU Bench V1: Success vs. Throughput](https://browser-use.com/images/benchmark/accuracy_vs_throughput_dark.png)

## The model breakdown

### Browser Use Cloud (`bu-ultra`) -- 78.0%

The clear winner. Not just the highest accuracy but also the fastest at ~14 tasks per hour. Each step is slower than a raw frontier LLM call, but bu-ultra completes tasks in far fewer steps, so total wall-clock time is lower.

This isn't just a better model. [Introduced in Browser Use 1.0](https://browser-use.com/posts/speed-matters), it's a purpose-built agent with [stealth browser infrastructure](https://browser-use.com/posts/stealth-benchmark), CAPTCHA solving, persistent filesystem, and optimized tool orchestration. The 16-point gap over the best frontier model comes from full-stack optimization.

**Verdict**: Best performance, highest throughput, zero setup. Use this.

### ChatBrowserUse-2 (OSS + Cloud LLM) -- 63.3%

Our model specifically optimized for browser automation, running on the open-source library. It outperforms every standalone frontier model while being faster and cheaper per task.

**Verdict**: Best option if you need custom tools or self-hosting but still want top-tier accuracy.

### Claude -- 62.0% (opus), 59.0% (sonnet)

Claude models are the strongest standalone frontier option. Claude-opus-4-6 leads all non-Browser Use models at 62%. When the agent needs to execute custom JavaScript or extract complex structured data, Claude is unparalleled.

Claude-sonnet-4-6 at 59% is close behind opus at roughly half the cost, making it the best value among Anthropic models.

**Verdict**: Best standalone frontier model. Use Claude when your workflow relies on custom code execution.

### Gemini -- 59.3% (3-1-pro), 35.2% (2.5-flash)

Gemini-3-1-pro scores 59.3%, neck and neck with Claude Sonnet. Strong vision, low latency, and massive context windows that handle enormous DOMs without issue.

Gemini-2.5-flash at 35.2% is the fastest cheap option but accuracy drops hard. You get what you pay for.

**Verdict**: Gemini-3-1-pro is a strong alternative to Claude. Flash is only for cost-sensitive, low-stakes tasks.

### OpenAI -- 52.4% (gpt-5), 37.0% (gpt-5-mini)

GPT-5 scores 52.4% and is the slowest model on the benchmark at ~6 tasks per hour. GPT-5-mini at 37% doesn't make up for the speed gap.

Recent OpenAI models have not kept pace with Claude and Gemini on browser automation tasks.

**Verdict**: Falling behind competitors. Use Claude or Gemini instead.

## Recommendations

- **Want the best agent?** Use [Browser Use Cloud](https://cloud.browser-use.com) (`bu-ultra`). 78% accuracy, fastest throughput, no setup.
- **Need custom tools or self-hosting?** Use the open-source library with `ChatBrowserUse-2`. 63.3%, still beats every frontier model.
- **Prefer a standalone frontier LLM?** Claude-opus-4-6 (62%) or gemini-3-1-pro (59.3%).
- **On a budget?** Claude-sonnet-4-6 (59%) gives near-opus accuracy at lower cost.

The benchmark is open source at [github.com/browser-use/benchmark](https://github.com/browser-use/benchmark). If you're an LLM provider looking to test at scale, reach out at support@browser-use.com.