GPT-6 Astra browser use: we tested ego-browser, agent-browser and playwright-cli on 28 web tasks

We would recommend ego-browser if you run GPT-6 Astra as a browser agent. GPT-6 Astra finishes most web tasks whichever tool it drives, so the choice comes down to how long you wait and what you pay.
In this test, we ran GPT-6 Astra on 28 web tasks, with one attempt per task for each tool. ego-browser and playwright-cli each passed 23 tasks (82.1%). agent-browser passed 20 (71.4%). The bigger difference was speed. On the 19 tasks all three tools completed, ego-browser averaged 263 seconds per task, compared with 332 for agent-browser and 376 for playwright-cli. Costs were much closer between ego-browser and playwright-cli: $2.22 versus $2.26 per task. agent-browser averaged $2.64 on those same tasks.
With ego-browser, Astra often handled work across two sites in a single short script. It needed fewer calls to finish, which helps explain the shorter runtimes. Each call cost a little more, though, leaving its cost per task close to playwright-cli’s.

Success rate
In our previous article, we ran the same 28 tasks with Grok 4.7. ego-browser had a clear edge as it completed more tasks, while agent-browser and playwright-cli repeatedly ran out of steps without producing an answer. However, as Astra generally needed fewer steps than Grok 4.7, the gap on success rate narrows. With ego-browser and playwright-cli completing nearly identical sets of tasks, speed became the more discerning difference between them.
Taking a closer look at the results, ego-browser and playwright-cli finished almost exactly the same tasks. Their results differed on just two, with each completing one that the other missed. agent-browser had more trouble finishing. It failed eight tasks, reaching the step limit on three: the flight search and two lengthy jobs collecting information across multiple sites. Astra completed both collection jobs with ego-browser and playwright-cli.
Speed
ego-browser is the faster tool on almost every task, so its lead isn't an average carried by a few tasks. It beat playwright-cli on 17 of the 19 shared tasks and agent-browser on 16. Per task, playwright-cli took 43% longer and agent-browser 26% longer.

The time is saved in the browser, not in the model. Time spent waiting on GPT-6 Astra was similar with all three tools, between 133 and 160 seconds per task. Time spent running browser commands was 94 seconds with ego-browser, 138 seconds with agent-browser and 203 seconds with playwright-cli. ego-browser makes the fewest calls, and a playwright-cli call took about 6 seconds to run, against about 3.5 seconds with either of the other tools.
The gap is widest on long jobs. On the five multi-site collection jobs every tool passed, a job took 442 seconds with ego-browser, 584 seconds with agent-browser and 683 seconds with playwright-cli.
The two tasks where playwright-cli was quicker are short lookups on one or two sites: a check of four companies in the UK company register, and a browser-support check for six web features. There is little to batch on tasks like these, and playwright-cli finished them 69 and 43 seconds sooner.
Cost
Compared on the tasks they all past, ego-browser is the least expensive. A call cost about 8 cents with ego-browser and about 7 cents with the other two, as the former contains a short script. Even though each call sent out by ego-browser costs slightly more, fewer calls and higher success rate cancels out the extra charge. The more complicated your task is, the more money you will save with ego-browser.
The tools work differently
It's easier to see how these tools work differently if you look into the steps they take to finish a hard job. One task asks for the first 2026 rate decision at 12 central banks, taken from each bank's own site. This task is especially difficult since the agent needs to work across multiple sites and extract the exact number we ask for.

With ego-browser, GPT-6 Astra opened the Federal Reserve's statement in one page and the ECB's decision page in a second page, in the same script. Half of its browser calls on this job opened two pages. It reached the twelfth bank on call 34 and delivered a correct brief after 40 tool calls. The job took 414 seconds and cost $4.20.

This is ego(lite)'s Spaces view. Each agent task runs in a space of its own, apart from your tabs. The space outlined in blue is the agent's, and the others are workspaces you keep using as normal.
With playwright-cli it took the same route one page at a time. It reached the twelfth bank on call 64 and answered after 85 tool calls, 753 seconds and $6.48. With agent-browser it reached the twelfth bank on call 92, and the 100 steps ran out before it wrote the answer. That attempt cost $9.65.
The same habit shows across the whole test. Over all 28 tasks, 16% of GPT-6 Astra's ego-browser calls opened two or more pages, against 4% of its agent-browser calls and 6% of its playwright-cli calls.
Which tool to use with GPT-6 Astra
For GPT-6 Astra we'd use ego-browser. How much it matters depends on what you're optimizing for:
Getting tasks done. ego-browser and playwright-cli are equally good choices. We'd avoid agent-browser for long multi-site jobs.
Waiting less. ego-browser, and the longer the job, the more time it saves.
Spending less. ego-browser or playwright-cli. Cost won't settle the choice between those two.
Short lookups on one or two sites. Any of the three will do.
To get the setup behind these results, run gpt-6-astra with reasoning set to high in the pi agent, with bash and file tools. ego-browser comes with ego(lite) and drives ego(lite) as the agent's browser. agent-browser and playwright-cli drive Google Chrome.
FAQ
What is the best browser automation tool for GPT-6 Astra? ego-browser was the best fit for GPT-6 Astra in our benchmark, mainly on speed. On 28 web tasks run once per tool in the same pi agent, GPT-6 Astra passed 23 with ego-browser, 20 with agent-browser and 23 with playwright-cli. On the 19 tasks every tool passed, a task took 263 seconds with ego-browser, 332 seconds with agent-browser and 376 seconds with playwright-cli.
Is ego-browser better than playwright-cli for AI agents? With GPT-6 Astra, ego-browser was faster than playwright-cli and equal on success and cost. Both tools passed 23 of 28 tasks. On the 19 tasks every tool passed, ego-browser took 263 seconds and $2.22 per task, and playwright-cli took 376 seconds and $2.26.
Is ego-browser better than agent-browser? With GPT-6 Astra, ego-browser finished more tasks than agent-browser, and it was faster and cheaper. GPT-6 Astra passed 23 of 28 tasks with ego-browser and 20 of 28 with agent-browser. On the 19 tasks every tool passed, ego-browser took 263 seconds and $2.22 per task, and agent-browser took 332 seconds and $2.64.
Is agent-browser better than playwright-cli? With GPT-6 Astra, playwright-cli finished more tasks and cost less, and agent-browser was faster. GPT-6 Astra passed 23 of 28 tasks with playwright-cli and 20 of 28 with agent-browser. On the 19 tasks every tool passed, agent-browser took 332 seconds and $2.64 per task, and playwright-cli took 376 seconds and $2.26.
How good is GPT-6 Astra at browser use? GPT-6 Astra completed most of our web tasks with every browser tool we tried. It passed 23 of 28 tasks with ego-browser and with playwright-cli, and 20 of 28 with agent-browser. It passed all four business web app workflows and all three research tasks with every tool.
How much does it cost to run GPT-6 Astra as a browser agent? In our test, GPT-6 Astra cost between $2.22 and $2.64 per task on the 19 tasks every tool passed. It cost $2.22 with ego-browser, $2.64 with agent-browser and $2.26 with playwright-cli. Counting failed attempts, a pass cost $3.69 with ego-browser, $5.29 with agent-browser and $3.69 with playwright-cli.
How long does GPT-6 Astra take to finish a browser task? GPT-6 Astra took 263 to 376 seconds per task in our test, depending on the browser tool. On the 19 tasks every tool passed, a task took 263 seconds with ego-browser, 332 seconds with agent-browser and 376 seconds with playwright-cli. Multi-site collection jobs took longer: on the five every tool passed, a job took 442 seconds with ego-browser and 683 seconds with playwright-cli.
Which browser tool needs the fewest tool calls with GPT-6 Astra? ego-browser needed the fewest tool calls with GPT-6 Astra. On the 19 tasks every tool passed, GPT-6 Astra made 27.1 tool calls per task with ego-browser, 38.4 with agent-browser and 33.3 with playwright-cli. It writes a short script for each ego-browser call, and 16% of those calls opened two or more pages.
Can GPT-6 Astra collect data from many websites in one run? Yes, in most cases, and the browser tool affects how often it finishes. On 10 jobs that collect the same fields from 8 to 12 official sites, GPT-6 Astra passed 8 with ego-browser, 6 with agent-browser and 8 with playwright-cli. With agent-browser, two of those jobs ran out of the 100-step budget before giving an answer.
Which browser tool is best for web app automation with GPT-6 Astra? All three tools completed our web app workflows with GPT-6 Astra, and ego-browser was the quickest. On four business app tasks, such as matching invoices to payments, GPT-6 Astra passed every attempt with each tool. A task took 204 seconds and $1.21 with ego-browser, 236 seconds and $1.27 with agent-browser, and 258 seconds and $1.26 with playwright-cli.
How we tested
We ran one model, GPT-6 Astra (gpt-6-astra, reasoning set to high), in the pi agent 1.5.2 with bash and file tools. The browser tools were ego-browser v2.2.0 with ego(lite) 0.6.3.6, agent-browser 0.39.1 and playwright-cli 0.1.22. Each tool came with its own skill file. With ego-browser, the model reads pages by writing short scripts that run in ego(lite). With agent-browser and playwright-cli, it reads pages through each CLI's snapshot and eval commands in Google Chrome. All runs were on the same Linux machine on November 13, 2026.
The task set has 28 tasks in five types of work, and each task belongs to one type. Multi-site data collection has 10 jobs, such as rate decisions at 12 central banks. Social media verification has 3 tasks that match company posts to press releases. Search and compare has 8 tasks, such as an inflation brief from FRED and BLS. Research and synthesis has 3 tasks, such as tracing a Wikipedia table to its sources. Web app automation has 4 tasks, such as matching invoices in a test billing app.
Each tool ran each task once, so an attempt is one run of one task with one tool. An attempt passes only when every scored item in its final answer is correct. Each attempt had a budget of 100 steps, where a step is one reply from the model. A tool call is one call the model makes to a tool. Time is wall-clock time from the start of an attempt to its end, including time spent waiting on the model API. Cost is the API cost of the attempt in US dollars.
Tool calls, time and cost per task are compared on the 19 tasks every tool passed. Time waiting on the model and time running browser commands are the two parts of that time, averaged over the same 19 tasks. Cost per pass is a tool's total cost over all 28 attempts divided by its passes. The share of calls that opened two or more pages was counted over every browser call in all 28 attempts per tool.
Full results
Each cell shows whether the attempt passed, how long it took and what it cost. Each tool ran each task once. Time includes time spent waiting on the model API.
| Task | Type of work | ego-browser | agent-browser | playwright-cli |
|---|---|---|---|---|
| Compliance claims on 9 SaaS trust centers | Multi-site data collection | Pass · 447 s · $5.07 | Fail · 504 s · $4.09 | Pass · 884 s · $5.56 |
| Current and next adult minimum wage in 9 jurisdictions | Multi-site data collection | Pass · 506 s · $4.31 | Pass · 700 s · $7.34 | Pass · 615 s · $3.99 |
| First 2026 rate decision at 12 central banks | Multi-site data collection | Pass · 414 s · $4.20 | Fail · 695 s · $9.65 | Pass · 753 s · $6.48 |
| Free-tier allowances on 9 serverless platforms | Multi-site data collection | Pass · 258 s · $1.57 | Pass · 333 s · $1.76 | Pass · 401 s · $2.01 |
| One album's track count, runtime and label on 9 music services | Multi-site data collection | Fail · 464 s · $4.21 | Fail · 397 s · $4.19 | Fail · 375 s · $3.04 |
| Opening keynote details from 9 developer conferences | Multi-site data collection | Pass · 681 s · $6.94 | Fail · 850 s · $9.46 | Pass · 840 s · $7.60 |
| Parameters, context and license of 9 open-weight models | Multi-site data collection | Fail · 997 s · $6.26 | Pass · 1172 s · $8.86 | Fail · 1427 s · $4.57 |
| Range, battery and charging specs from 9 electric SUV makers | Multi-site data collection | Pass · 585 s · $7.60 | Pass · 696 s · $6.15 | Pass · 903 s · $7.38 |
| Weight, battery, resolution and burst rate from 8 camera makers | Multi-site data collection | Pass · 416 s · $4.40 | Pass · 447 s · $2.98 | Pass · 825 s · $3.69 |
| Weight, battery, screen and IP rating of 9 flagship phones | Multi-site data collection | Pass · 444 s · $4.35 | Pass · 747 s · $5.51 | Pass · 673 s · $4.50 |
| Match a week of launch posts to two Wikipedia launch lists | Social media verification | Fail · 402 s · $3.09 | Fail · 933 s · $5.14 | Fail · 668 s · $4.14 |
| Match product-launch posts by 8 companies to their press releases | Social media verification | Pass · 380 s · $3.58 | Fail · 437 s · $5.10 | Fail · 492 s · $4.12 |
| Match security-release posts by 8 open-source projects to their announcements | Social media verification | Fail · 274 s · $2.03 | Fail · 228 s · $1.45 | Fail · 267 s · $0.93 |
| Build an inflation brief from FRED series and check it on BLS | Search and compare | Pass · 99 s · $0.90 | Pass · 227 s · $1.78 | Pass · 156 s · $1.07 |
| Cheapest nonstop fares from three cities on a flight search site | Search and compare | Fail · 802 s · $7.33 | Fail · 840 s · $7.71 | Pass · 860 s · $5.43 |
| Go/no-go for six CSS and web features across three browsers | Search and compare | Pass · 230 s · $1.51 | Pass · 188 s · $1.28 | Pass · 187 s · $1.00 |
| KYC screen of four UK companies in the company register | Search and compare | Pass · 245 s · $1.62 | Pass · 198 s · $1.95 | Pass · 176 s · $0.96 |
| Price and spec three Mac laptops in one configuration | Search and compare | Pass · 285 s · $2.72 | Pass · 336 s · $4.50 | Pass · 388 s · $2.70 |
| Reconcile 10 clinical trials with their drug labels | Search and compare | Pass · 213 s · $1.76 | Pass · 307 s · $2.89 | Pass · 493 s · $3.19 |
| Reconcile box office figures for six films across two sites | Search and compare | Pass · 194 s · $1.34 | Pass · 226 s · $1.54 | Pass · 271 s · $1.48 |
| Turn a lecture video, its code and its paper into study notes | Search and compare | Pass · 111 s · $1.01 | Pass · 181 s · $1.47 | Pass · 149 s · $1.01 |
| Count models above EU compute thresholds in a public model database | Research and synthesis | Pass · 114 s · $0.62 | Pass · 125 s · $0.75 | Pass · 134 s · $0.84 |
| Read scores and per-task cost from an interactive leaderboard chart | Research and synthesis | Pass · 208 s · $2.07 | Pass · 279 s · $3.34 | Pass · 399 s · $2.80 |
| Trace 10 cells of a Wikipedia table to their cited sources | Research and synthesis | Pass · 270 s · $1.55 | Pass · 374 s · $1.88 | Pass · 344 s · $1.26 |
| Book a policy-compliant round trip, aisle seats and a hotel near the venue | Web app automation | Pass · 224 s · $1.50 | Pass · 249 s · $1.53 | Pass · 239 s · $1.72 |
| Edit, add and delete service monitors to match a monitoring standard | Web app automation | Pass · 164 s · $0.99 | Pass · 227 s · $1.13 | Pass · 276 s · $1.15 |
| Find items below reorder point and create purchase orders from a CSV | Web app automation | Pass · 182 s · $1.01 | Pass · 222 s · $0.99 | Pass · 193 s · $0.81 |
| Match February invoices to payments and dispute the unpaid ones | Web app automation | Pass · 247 s · $1.34 | Pass · 246 s · $1.44 | Pass · 326 s · $1.37 |



