Browser Use found the right price on 29 of 34 stores, and duplicate product names caused most misses

Browser Use is an open-source AI browser agent. You give it a task in plain words, and it clicks, types and reads pages in a real browser until it has an answer. It has more than 118,000 GitHub stars, and it got 8.1 million PyPI downloads in 30 days in my stars test.

Teams that collect product data have to choose between an agent like this and a script. So I gave Browser Use a job that data teams run every day. Go to a store, find a product by its name, and return its price.

On 34 real stores, it returned the right product and price for 29 in the first run and 28 in the second. It returned an answer in all 68 tasks. Most of its misses weren’t reading errors. The store had more than one product with the same name, and the agent found a different one.

The test#

I ran everything from Mumbai on October 18, 2026.

• Stores. 34 Shopify stores, built with the same rule as my October 16 price test. For each store, I took the first product with one price from the store’s public product list. I kept it only if the price in its JSON-LD also appeared in the visible page text. The October 16 store list wasn’t saved, so this is a new set.
• Agent. Browser Use 0.13.11 with its default settings, a local Chrome window, and OpenAI’s GPT-6 Luna at the library’s default reasoning effort. It ran one task at a time, with up to 30 steps each.
• Runs. I ran all 34 tasks twice.
• Grading. After the runs, I opened each product page and each page the agent returned in Chrome, from the same place. I compared the agent’s price with the price that page showed. I checked every answer that didn’t match by hand, against the agent’s saved step log.

The prompt for each store was this one, with the store’s homepage and the product name filled in.

Go to [homepage]. Find the product named “[product name]” on this store. Open its product page and return the price exactly as that page shows it, the currency, and the product page URL.

The agent returned the price, the currency and the URL as structured fields.

Some stores showed prices in Indian rupees, because the test ran from India. I graded each answer against the currency its page showed.

Browser Use 0.13.11 didn’t recognize GPT-6 Luna as a reasoning model, so it sent a temperature setting that the model rejects. I added the model to the library’s list of reasoning models, which is an option of its OpenAI client.

Results#

33 of the 34 stores got the same result in both runs. Each task took a median of 55 seconds and about 7 steps.

The same name on more than one page#

On 4 stores, the agent opened a real product page with the exact name I gave it, and returned that page’s price correctly. It just wasn’t the page in my list. It happened on the same 4 stores in both runs.

• Stanley has 3 product pages called “The IceFlow™ Flip Straw Tumbler | 30 OZ”. They cost $40, $35 and $26.25. The agent opened the $26.25 page in the first run and the $35 page in the second.
• NoBull has a page for each color, and every one is called “Women’s Outwork Flex”.
• Outdoor Voices has a page for each color of its “CloudKnit Cami”. The color the agent opened cost $23.97, and the one in my list cost $34.
• Outerknown sells used items under its Outerworn label. Each used item has its own page, and several share a name.

A product name isn’t an ID. If your agent or pipeline gets a product by its name, it can return a real price for a different product. Give it a URL or a SKU when you have one.

The real mistakes#

The agent made 3 mistakes in 68 tasks.

A per-pair price for a 2-pack, in both runs. On True Classic, the largest price on the page is the price per pair. The price of the 2-pack is smaller, on the right. The agent returned the per-pair price both times. The simplest rule in my October 16 test made the same mistake on a 2-pack.

The Canadian store, once. On Herschel, the agent didn’t find the product in the store’s own search results. It then searched DuckDuckGo, clicked a result on herschel.ca and returned its price in Canadian dollars. The task said “on this store”. In the first run, the agent found the right page on herschel.com.

The agent left the store for a search engine in 4 tasks in each run. 7 of those 8 tasks still ended on the right page. If your task must stay on one site, check the domain of the URL the agent returns.

Time and cost#

Each agent task used a median of about 120,000 input tokens and about 1,500 output tokens. At OpenAI’s list price for GPT-6 Luna on October 18, 2026, $0.10 per million input tokens and $0.50 per million output tokens, each run of 34 tasks cost about $0.45. That’s about $0.013 per product, or about $13 per 1,000. Browser Use’s usage report showed no cached tokens, so the real cost may be lower.

For comparison, I fetched each known product URL once and read the price from its JSON-LD. That took a median of 1.0 second and found all 34 prices. But I chose these stores because they publish that data, and I already had each URL. Finding the URL is the part the agent did.

Browser Use also advertises a faster mode. It was on a waitlist when I tested, so I didn’t test it.

A warning from the first try#

Before the full test, I ran 5 tasks with an older model, GPT-5.6 Luna. 3 of the 5 answers came back with an empty price, but Browser Use marked all 5 as successful. When I ran one of them again, it returned the right price.

5 tasks is too few to compare models. But it shows why the agent’s own success flag isn’t enough. Check that the fields you need have values.

What production teams can do with this#

At Extract Summit 2026, Zyte described a setup where AI writes the extraction code, and AI also extracts data directly from a sample of pages to check that code. The code does the repeated work. This test shows why that split makes sense.

• Use the agent where a script needs a human first. Finding a product, a page or a URL on a site you haven’t seen is where the agent saves real time.
• Give it IDs, not names. A URL or a SKU removes the most common miss in this test.
• Check what it returns. Check the domain, the currency, and the pack size. Don’t rely on the agent’s success flag alone.
• Run repeated collection as a script. At about a minute per product, 1,000 products take more than 15 hours one at a time. A script that already has the URLs reads them in minutes.

All 34 stores#

If you want the step logs, the results or the scripts, email me and I’ll share them.