We benchmarked Grok 4.7 browser use with ego-browser, agent-browser and playwright-cli

We would recommend ego-browser if you run Grok 4.7 as a browser agent. It finishes more tasks than agent-browser or playwright-cli, and on the work all three can do, it's the fastest and the cheapest.
Because ego-browser is designed in a cost-efficient way, Grok 4.7 writes a short script for each call that opens a page, clicks and reads the result in one go. With the other two tools, each call does less. Every task in our test had a budget of 100 steps, and a tool that gets more done per step leaves the model enough room to reach the end of a long task without exceeding the budget.
The lead isn't the same size everywhere. Its advantage is largest on long tasks spanning multiple sites and smaller on short web app workflows. These tasks required fewer steps, and all three tools completed them without a problem.

Success rate
What separates the three tools is whether they can reach an answer before using up the budget. Most failures in this test were "works in progress". The model was still working when its 100 steps ran out, and its last message was a progress note. That happened 16 times with playwright-cli and 13 times with agent-browser, which accounts for most of the failures those two tools had. With ego-browser it happened 6 times.
Head to head, ego-browser passed 7 tasks that playwright-cli failed and 5 that agent-browser failed, and in every one of those cases the other tool had run out of steps. We would interpret it as a sign that ego-browser approaches the same problem more efficiently.
Speed
The step count is where the tools differ most. An ego-browser call ran 2.5 browser operations on average. A playwright-cli call is close to a single command. Grok 4.7 does chain commands with agent-browser, about 2.3 per call, but roughly a quarter of those are waits. As a result, ego-browser needed the fewest calls on every one of the nine tasks all three tools passed, about half as many as playwright-cli.
Cost
On the tasks all three tools passed, ego-browser was the cheapest on every one. The same work cost about 1.4 times as much with agent-browser and 1.8 times as much with playwright-cli.
There is one complication. As Grok 4.7 has to write the script it runs, ego-browser cost more for a single call. However, tasks getting finished with far fewer calls more than enough pays for that. Dividing everything a tool spent across all 28 tasks by its passes gives $5.28 per pass for ego-browser, against $7.37 for agent-browser and $7.24 for playwright-cli.
Where the gap shows up
Multi-site data collection
These are the toughest jobs in the set. Each asks for the same fields from 8 to 12 official sites, such as the first 2026 rate decision at 12 central banks. With agent-browser and playwright-cli, Grok 4.7 finished none of the ten. With ego-browser it finished three: the central bank decisions, a camera spec comparison and a phone spec comparison.
Even those three passes used between 80 and 104 tool calls each, right at the edge of the budget, and three out of ten isn't reliable. The other seven jobs failed with every tool, and in five of them Grok 4.7 ran out of steps with ego-browser too. If this is your main workload, we'd also look at raising the step budget or splitting the jobs, neither of which we tested here.

Search and compare
In an inflation brief, Grok 4.7 had to pull three monthly series from FRED, compute the December-over-December changes, and check the series and their 2025 annual averages on the BLS site.
With all three tools it took the same route and got stuck at the same point: whether BLS publishes an annual average for egg prices. With ego-browser it reached that question after 22 calls, worked out that BLS doesn't publish one, and delivered a complete brief at call 52. With agent-browser and playwright-cli it arrived at the same question after roughly 36 and 40 calls, and was still searching BLS pages when the budget ran out.
The same pattern shows over and over again with ego-browser passing the task with fewer steps.

Web app automation
Every tool passed all four of these short workflows inside a business web app, but ego-browser cost less tokens. In one of them, Grok 4.7 had to find the items below their reorder point in a test inventory app and raise one purchase order per supplier from a CSV import. It did that in 36 calls for $0.69 with ego-browser, 50 calls for $0.84 with agent-browser, and 70 calls for $1.34 with playwright-cli. The other three workflows followed the same order, and ego-browser was also the quickest on all four.
Which tool to use with Grok 4.7
For most browser work with Grok 4.7, we'd use ego-browser. How much you gain depends on the work:
Long, multi-site jobs. The choice matters most here. ego-browser is the only tool that gives Grok 4.7 a chance of finishing.
Search and comparison across a few sites. ego-browser passes more tasks.
Short workflows in a web app. Any of the three will do the job, and ego-browser does it for less.
Whichever tool you pick, review the answers.
To get the setup behind these results, run grok-4.7 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 Grok 4.7? ego-browser was the best fit for Grok 4.7 in our benchmark. On 28 real web tasks, run once per tool in the same pi agent, Grok 4.7 passed 57.1% of attempts with ego-browser, 42.9% with agent-browser and 39.3% with playwright-cli. The gain came from long tasks, where the other tools ran out of the 100-step budget more often. On simple business web app workflows, all three tools passed every task. Full setup and per-task results are in How we tested.
Is ego-browser better than agent-browser for AI agents? With Grok 4.7, ego-browser finished more tasks and cost less on shared work. Grok 4.7 passed 16 of 28 attempts with ego-browser and 12 of 28 with agent-browser, on the same 28 tasks and the same agent. On the 9 tasks both passed, ego-browser cost $1.01 per task and agent-browser $1.43. The difference is the number of calls: 33.4 per task with ego-browser and 53.4 with agent-browser. Full setup and per-task results are in How we tested.
Is agent-browser better than playwright-cli? With Grok 4.7 the two were close on success and apart on cost. Grok 4.7 passed 12 of 28 attempts with agent-browser and 11 of 28 with playwright-cli, on the same 28 tasks. On the 9 tasks all tools passed, agent-browser cost $1.43 per task and playwright-cli $1.86. Grok 4.7 chained several actions into one agent-browser call, so it made fewer calls than with playwright-cli. Full setup and per-task results are in How we tested.
How good is Grok 4.7 at browser use? Grok 4.7 handles short browser tasks well and struggles with long multi-site jobs. It passed every business web app task with all three browser tools we tried, 4 of 4 each. On jobs that collect the same fields from 8 to 12 sites, it passed 3 of 10 attempts with ego-browser and none with agent-browser or playwright-cli. Most of its failures came from running out of the 100-step budget. Full setup and per-task results are in How we tested.
Why does Grok 4.7 run out of steps on browser tasks? Grok 4.7 runs out of steps when each step does too little for a long task. With a 100-step budget, it ran out of steps in 16 of 28 attempts with playwright-cli, 13 with agent-browser and 6 with ego-browser. With ego-browser it needed fewer calls, 33.4 per task against 53.4 with agent-browser and 68.1 with playwright-cli on tasks all tools finished, so the same budget covered more of the task. Most of these failures came from jobs that span 8 or more sites. Full setup and per-task results are in How we tested.
How much does it cost to run Grok 4.7 as a browser agent? With ego-browser, Grok 4.7 cost $1.01 per task on the tasks every tool finished. With agent-browser it cost $1.43 and with playwright-cli $1.86, on the same 9 tasks. Counting failed attempts, a pass cost $5.28 with ego-browser, $7.37 with agent-browser and $7.24 with playwright-cli. Attempts that ran out of steps cost the most, $3.57 to $6.12 on average depending on the tool. Full setup and per-task results are in How we tested.
Which browser tool uses the fewest tokens with Grok 4.7? playwright-cli used the fewest output tokens and ego-browser the most, but ego-browser needed far fewer calls. Over 28 attempts each, Grok 4.7 produced 617,992 output tokens with playwright-cli, 714,782 with agent-browser and 960,733 with ego-browser, because it writes a script for each ego-browser call. On tasks all tools finished, that still made ego-browser the cheapest, at $1.01 per task. Full setup and per-task results are in How we tested.
Can Grok 4.7 scrape data from many websites in one run? Not reliably, with any of the browser tools we tested. On 10 jobs that collect the same fields from 8 to 12 official sites, Grok 4.7 passed 3 with ego-browser and none with agent-browser or playwright-cli. Most attempts ran out of the 100-step budget before giving an answer. ego-browser's three passes came from needing fewer calls, and they still used 80 to 104 tool calls each. Full setup and per-task results are in How we tested.
Which browser tool is best for web app automation with Grok 4.7? All three tools completed our web app workflows with Grok 4.7, so cost decides. On four business app tasks, such as matching invoices to payments and raising purchase orders, Grok 4.7 passed every attempt with ego-browser, agent-browser and playwright-cli. ego-browser cost $0.84 per task, agent-browser $1.21 and playwright-cli $1.69. Full setup and per-task results are in How we tested.
How we tested
We ran one model, Grok 4.7 (grok-4.7 on the xAI API, 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 9 tasks every tool passed. Cost per pass is a tool's total cost over all 28 attempts divided by its passes. Commands per call were counted over all 28 attempts per tool. For ego-browser, each browser operation in a script counts, except checks of the page address or title.
Full results
Each cell shows whether the attempt passed and what it cost. Each tool ran each task once.
| Task | Type of work | ego-browser | agent-browser | playwright-cli |
|---|---|---|---|---|
| Compliance claims on 9 SaaS trust centers | Multi-site data collection | Fail · $5.68 | Fail · $3.39 | Fail · $2.67 |
| Current and next adult minimum wage in 9 jurisdictions | Multi-site data collection | Fail · $3.65 | Fail · $3.66 | Fail · $2.47 |
| First 2026 rate decision at 12 central banks | Multi-site data collection | Pass · $4.65 | Fail · $4.80 | Fail · $2.77 |
| Free-tier allowances on 9 serverless platforms | Multi-site data collection | Fail · $6.74 | Fail · $5.05 | Fail · $4.08 |
| One album's track count, runtime and label on 9 music services | Multi-site data collection | Fail · $2.78 | Fail · $3.46 | Fail · $3.33 |
| Opening keynote details from 9 developer conferences | Multi-site data collection | Fail · $5.16 | Fail · $4.84 | Fail · $3.87 |
| Parameters, context and license of 9 open-weight models | Multi-site data collection | Fail · $12.00 | Fail · $6.39 | Fail · $7.27 |
| Range, battery and charging specs from 9 electric SUV makers | Multi-site data collection | Fail · $4.94 | Fail · $4.76 | Fail · $3.04 |
| Weight, battery, resolution and burst rate from 8 camera makers | Multi-site data collection | Pass · $4.21 | Fail · $4.96 | Fail · $2.77 |
| Weight, battery, screen and IP rating of 9 flagship phones | Multi-site data collection | Pass · $5.15 | Fail · $4.82 | Fail · $4.07 |
| Match a week of launch posts to two Wikipedia launch lists | Social media verification | Pass · $0.52 | Pass · $0.89 | Pass · $1.51 |
| Match product-launch posts by 8 companies to their press releases | Social media verification | Fail · $3.14 | Fail · $4.39 | Fail · $3.58 |
| Match security-release posts by 8 open-source projects to their announcements | Social media verification | Fail · $4.25 | Fail · $4.16 | Fail · $3.82 |
| Build an inflation brief from FRED series and check it on BLS | Search and compare | Pass · $1.72 | Fail · $4.67 | Fail · $2.93 |
| Cheapest nonstop fares from three cities on a flight search site | Search and compare | Pass · $3.37 | Fail · $5.02 | Fail · $3.69 |
| Go/no-go for six CSS and web features across three browsers | Search and compare | Fail · $1.10 | Fail · $2.27 | Pass · $2.10 |
| KYC screen of four UK companies in the company register | Search and compare | Fail · $1.11 | Pass · $1.71 | Pass · $1.26 |
| Price and spec three Mac laptops in one configuration | Search and compare | Pass · $3.26 | Pass · $3.54 | Fail · $4.06 |
| Reconcile 10 clinical trials with their drug labels | Search and compare | Pass · $1.47 | Pass · $1.96 | Fail · $2.72 |
| Reconcile box office figures for six films across two sites | Search and compare | Pass · $0.84 | Pass · $1.66 | Pass · $1.77 |
| Turn a lecture video, its code and its paper into study notes | Search and compare | Pass · $1.44 | Pass · $1.51 | Pass · $1.53 |
| Count models above EU compute thresholds in a public model database | Research and synthesis | Pass · $0.54 | Pass · $1.13 | Pass · $0.96 |
| Read scores and per-task cost from an interactive leaderboard chart | Research and synthesis | Pass · $2.38 | Pass · $2.79 | Pass · $4.18 |
| Trace 10 cells of a Wikipedia table to their cited sources | Research and synthesis | Fail · $1.02 | Fail · $1.72 | Fail · $2.40 |
| Book a policy-compliant round trip, aisle seats and a hotel near the venue | Web app automation | Pass · $0.98 | Pass · $1.54 | Pass · $2.11 |
| Edit, add and delete service monitors to match a monitoring standard | Web app automation | Pass · $0.69 | Pass · $0.73 | Pass · $1.40 |
| Find items below reorder point and create purchase orders from a CSV | Web app automation | Pass · $0.69 | Pass · $0.84 | Pass · $1.34 |
| Match February invoices to payments and dispute the unpaid ones | Web app automation | Pass · $1.01 | Pass · $1.75 | Pass · $1.90 |


