Claude Opus 5 Browser Use Benchmark: Which Tool to Choose?

Across 27 real web tasks, Opus 5 finished more of them with ego-browser than with any other tool we tested, and it finished them faster and for less. So ego-browser is the browser tool we'd pair with Claude Opus 5. With ego-browser it passed 24 of 27 tasks. Chrome DevTools CLI and Playwright CLI reached 18, Browser Use CLI 17 and agent-browser 15. On the 13 tasks every tool finished, ego-browser averaged 324.6 seconds and $1.702 per task. Chrome DevTools CLI came closest at 369.5 seconds and $1.845, then Browser Use CLI at 407.7 seconds and $1.975, Playwright CLI at 495.8 seconds and $2.052, and agent-browser at 543.6 seconds and $2.368.

The tools pull apart most on long work across many sites. We gave each tool 11 tasks that collect the same fields from 8 to 16 official sites. Opus 5 finished 8 of them with ego-browser and only 1 to 3 with the others. Most of those misses weren't wrong answers. Opus 5 ran out of its 100-step budget before writing anything, 7 to 9 times with the other tools and once with ego-browser.

We found the difference in how many tool calls the same work took. On the shared tasks, Opus 5 made 30.3 calls per task with ego-browser and 41.2 to 53.7 with the others. Against Chrome DevTools CLI, the closest competitor, its model time was level, 222.7 seconds per task against 222.4, so the time saved came from the browser side.

The same pattern shows up in the bill. The model and its price stay the same across tools, so any cost gap is usage. Fewer calls meant fewer cache reads, 1.32 million cached tokens per task against 1.69 million to 2.42 million. Against Chrome DevTools CLI, that came to 44.9 seconds and 14.3 cents saved per task.

Success rate

With ego-browser, Opus 5 finished the most tasks: 24 of 27 attempts, 88.9%. Chrome DevTools CLI and Playwright CLI reached 66.7% (18 of 27), Browser Use CLI 63.0% (17 of 27) and agent-browser 55.6% (15 of 27). All five passed every business web app task. The gap opens on long tasks that span many sites.

Most of the other tools' misses weren't wrong answers. Opus 5 ran out of steps before writing anything 7 times with Chrome DevTools CLI, 8 with Playwright CLI, 7 with Browser Use CLI and 11 with agent-browser. With ego-browser that happened once. Wrong values were rare with every tool: 2 with ego-browser, 2 with Chrome DevTools CLI, 1 with Playwright CLI, 3 with Browser Use CLI and 1 with agent-browser.

The central bank task shows the pattern. Opus 5 had to find 12 central banks' first scheduled rate decisions of 2026 on each bank's own site, with the date, the move, the policy rate, dissents and the next meeting. With ego-browser it went from the Fed through to the Reserve Bank of India and returned all 12 rows correctly in 86 tool calls. With each of the other four it ran out of steps before writing an answer.

Opus 5 with ego-browser reads the 12th bank's decision

STEP 74 OF 75 · RBI PRESS RELEASE

74goto('…?prid=62169')

Tool callego-browser · JS

ego-browser nodejs <<'EOF'const task = await taskSpace("…");const page = task.page("p1");await page.goto("https://www.rbi.org.in/scripts/BS_PressReleaseDisplay.aspx?prid=62169");await page.waitForTimeout(2500);const txt = await page.evaluate(() => document.body.innerText);const i=txt.indexOf("Monetary Policy Statement");console.log(txt.slice(i, i+3000));EOF

Opens RBI press release 62169 and prints 3,000 characters from the statement heading.

Outputexcerpt · 3,000 chars

Monetary Policy Statement, 2025-26 Resolution of the Monetary Policy Committee February 4 to 6, 2026…2. After a detailed assessment of the evolving macroeconomic and financial developments and the outlook, the MPC voted unanimously to keep the policy repo rate under the liquidity adjustment facility (LAF) unchanged at 5.25 per cent. Consequently, the standing deposit facility (SDF) rate remains at 5.00 per cent and the marginal standing facility (MSF) rate and the Bank Rate remains at 5.50 per cent. The MPC also decided to continue with the neutral stance.…

“Voted unanimously” and “unchanged at 5.25 per cent” become row 12: a hold at 5.25 with no dissents.

Final table · Row 12 · Reserve Bank of India

ego-browser2026-02-06 · Hold, 0 bp · 5.25 · 0 dissents (6–0) · next 2026-04-08Matches key

Chrome DevTools CLINo final answer, all 100 steps usedNo answer

Playwright CLINo final answer, all 100 steps usedNo answer

Browser Use CLINo final answer, all 100 steps usedNo answer

agent-browserNo final answer, all 100 steps usedNo answer

With ego-browser, Opus 5 missed 3 tasks. No tool finished the developer conference task. On the other two, Opus 5 returned a wrong value. In the EV task its own table had the Hyundai IONIQ 5 charging from 10 to 80% faster than the Porsche Macan Electric, yet its summary named the Macan the quickest.

ego-browser nodejs <<'EOF'const task = await taskSpace("…");const page = task.page("p1");await page.keyboard.press("Escape");await page.waitForTimeout(1000);let open = await page.evaluate(()=>!!document.querySelector('.zip-modal.modal-open'));console.log("modalOpen after esc:", open);if(open){ await page.evaluate(()=>{ document.querySelectorAll('.zip-modal,.backdrop').forEach(e=>e.style.display='none'); }); }await page.click("text=Expand all");await page.waitForTimeout(3000);const txt = await page.evaluate(() => document.body.innerText.replace(/\n{3,}/g,"\n\n"));const i = txt.indexOf("Powertrain & Handling");console.log(txt.slice(i, i+4000));EOF

…DC Fast Charging (10~80%) NACS 150 kW (400V)24 minutes30 minutes30 minutesDC Fast Charging (10~80%) w/CCS Adapter 50 kW1 hour 1 minute1 hour 21 minutes1 hour 21 minutesDC Fast Charging (10~80%) w/CCS Adapter >250 kW (800V)20 minutes20 minutes20 minutes…

Speed

ego-browser gave Opus 5 the lowest average time. On the 13 tasks all five tools passed, it averaged 324.6 seconds per task. Chrome DevTools CLI took 369.5, Browser Use CLI 407.7, Playwright CLI 495.8 and agent-browser 543.6. Head to head, ego-browser was faster than Chrome DevTools CLI on 8 of the 13 tasks, Playwright CLI and Browser Use CLI on 10 each, and agent-browser on 11.

Chrome DevTools CLI was the closest. It was faster on 5 of the 13, including the travel booking and monitor tasks. Playwright CLI's average includes one 2,127-second run on the SpaceX task, which accounts for most of its gap.

The time ego-browser saved goes with fewer tool calls, 30.3 per task against 41.2 to 53.7. Against Chrome DevTools CLI it was all browser time: model time was level, 222.7 seconds per task against 222.4, and tool time was 100.3 against 147.6. Against the other three, ego-browser also spent less time waiting on the model. Model time and tool time are timed separately and can overlap, so they don't add up exactly to the total.

The FRED and BLS task shows the gap on a job every tool finished. Opus 5 had to pull three monthly series from FRED, compute six years of December-over-December changes, then check the series IDs and 2025 averages on BLS. With ego-browser it finished in 24 tool calls and 181 seconds. The other four took 36 to 46 calls and 294 to 395 seconds, and all five got every value right.

Cost

ego-browser had the lowest average cost. On the 13 shared tasks it cost $1.702 per task, against $1.845 for Chrome DevTools CLI, $1.975 for Browser Use CLI, $2.052 for Playwright CLI and $2.368 for agent-browser. Over all 27 attempts, finished or not, ego-browser and Playwright CLI were close at $3.280 and $3.330. Chrome DevTools CLI averaged $3.401, Browser Use CLI $3.522 and agent-browser $3.739.

The model is the same in every tool, so the price is the same and the whole gap is usage. Cache reads are where it shows. Each model turn re-sends the conversation so far, and most of it is billed as cache reads. With ego-browser, Opus 5 read 1.32 million cached tokens per task, against 1.69 million to 2.42 million with the others.

At Anthropic's listed $0.50 per million cache-read tokens for Opus 5, that comes to $0.661 per task with ego-browser and $0.846 to $1.210 with the others. Output and cache writes don't explain the ranking. Output ran 16,047 to 20,927 tokens per task, lowest with Chrome DevTools CLI, and cache writes 59,662 to 63,386.

Multi-site data collection

If your agent collects the same fields from many official sites into one table, use ego-browser with Opus 5. On these 11 tasks, each spanning 8 to 16 sites, ego-browser finished 8. Playwright CLI finished 3, Chrome DevTools CLI and Browser Use CLI 2 each, and agent-browser 1.

Opus 5 averaged 73.4 tool calls per attempt here with ego-browser and 91.5 to 97.0 with the others. It ran out of steps before answering once with ego-browser, 7 times with Chrome DevTools CLI and Browser Use CLI, 8 with Playwright CLI and 9 with agent-browser.

No task in this group was finished by all five, so time and cost can't be compared task by task. Averaged over every attempt, finished or not, ego-browser took 919.1 seconds and $5.434 per task. The others took 952.0 to 1,112.2 seconds and $4.940 to $5.531, with Playwright CLI the lowest on cost.

The phone task is typical. Opus 5 had to pull weight, battery, display size, IP codes and the security-update promise for nine flagships, using only each brand's UK product pages. With ego-browser it worked through Apple, Samsung and Google down to Fairphone and returned all nine rows correctly in 74 tool calls. With each of the other four it ran out of steps before answering.

For jobs like these, ego-browser is the tool we'd pair with Opus 5: it finished 8 of 11 with fewer tool calls per attempt.

Cross-site lookups

If your agent checks figures across one to three sites, such as a registry against a drug label or a dataset against its source, ego-browser is the pick: it had the lowest average time. It passed all 9 of these tasks, as did Chrome DevTools CLI and Browser Use CLI. Playwright CLI passed 8 and agent-browser 7.

On the 7 tasks all five passed, ego-browser averaged 330.3 seconds per task. Chrome DevTools CLI took 385.8, agent-browser 416.1, Browser Use CLI 441.8 and Playwright CLI 636.9, including its 2,127-second SpaceX run. ego-browser had the lowest time on 5 of the 7 and beat each other tool on 5 or 6. It also had the lowest average cost here, $1.693 per task against $1.812 to $2.213.

The lecture task shows the shape of this work. Opus 5 had to turn Andrej Karpathy's GPT lecture into study notes, matching chapter times and losses on YouTube against the training script on GitHub and the paper's settings on arXiv. With ego-browser it finished in 35 tool calls and 245 seconds. Chrome DevTools CLI needed 66 calls and 426 seconds, Browser Use CLI 52 and 466, and Playwright CLI 70 and 696. agent-browser ran out of steps.

Opus 5 with ego-browser reads the training settings in one call

STEP 19 OF 21 · GITHUB GPT.PY

19goto('…/gpt.py')

Tool callego-browser · JS

cd /tmp && ego-browser nodejs <<'EOF'const task = await taskSpace("…");const page = task.page("p1");await page.goto("https://github.com/karpathy/ng-video-lecture/blob/master/gpt.py", { timeout: 60000 });await page.waitForTimeout(3000);const txt = await page.evaluate(() => {  const el = document.querySelector("#read-only-cursor-text-area") || document.querySelector("textarea");  if (el) return el.value;  const t = document.querySelector("table.highlight");  return t ? t.innerText : document.body.innerText.slice(0,3000);});console.log(txt.split("\n").slice(0,30).join("\n"));EOF

Reads the file from GitHub's code text area rather than the rendered lines, and prints the first 30 lines.

Outputexcerpt · 1,009 chars

…# hyperparametersbatch_size = 64 # how many independent sequences will we process in parallel?block_size = 256 # what is the maximum context length for predictions?max_iters = 5000eval_interval = 500learning_rate = 3e-4device = 'cuda' if torch.cuda.is_available() else 'cpu'eval_iters = 200n_embd = 384n_head = 6n_layer = 6dropout = 0.2…

Opus 5 checks each setting against what Karpathy says aloud: six match, and he never states the learning rate.

Final table · Script vs lecture · Embedding dimension, dropout

ego-browsern_embd = 384 · dropout = 0.2Matches key

Chrome DevTools CLIn_embd = 384 · dropout = 0.2Matches key

Playwright CLIn_embd = 384 · dropout = 0.2Matches key

Browser Use CLIn_embd = 384 · dropout = 0.2Matches key

agent-browserNo final answer, all 100 steps usedNo answer

For lookups like these, Chrome DevTools CLI and Browser Use CLI matched ego-browser's pass rate, and ego-browser got there sooner and for less on average.

Which browser tool to use with Claude Opus 5

Pair Claude Opus 5 with ego-browser. It finished the most tasks and had the lowest average time and cost on the tasks every tool finished. Its lead was largest on multi-site tasks.

By type of work, ego-browser's lead is widest on multi-site data collection and holds on cross-site lookups. On research and business web app tasks the five were closer. Browser Use CLI had the lowest time and cost on the two research tasks all five passed, and every tool passed all four web app tasks.

If getting the job done matters most, pick ego-browser, especially for agents that collect data from many sites. It finished 8 of 11 such tasks where the others finished 1 to 3. If speed matters most, ego-browser had the lowest average time, with Chrome DevTools CLI next. If cost matters most, ego-browser had the lowest average cost on the shared tasks, with Chrome DevTools CLI next.

If your agent checks figures across one to three sites, ego-browser was fastest and cheapest on average, and Chrome DevTools CLI and Browser Use CLI passed just as often. Among the other four, Chrome DevTools CLI was the closest to ego-browser on time and cost and tied for second on success rate.

We'd pick ego-browser for long multi-site jobs above all. In our runs it used fewer tool calls on those jobs and more often delivered an answer within 100 steps.

To get these results, run Claude Opus 5 (claude-opus-5) at high reasoning effort in Claude Code, with the Bash, Read, Write, Edit and Skill tools and a skill file for the browser tool. ego-browser comes with ego(lite) and drives it as the agent's browser. The other four drive Google Chrome.

FAQ

What is the best browser automation tool for Claude Opus 5?

ego-browser, in our benchmark. Across 27 web tasks in Claude Code with the same settings, Claude Opus 5 finished 88.9% with ego-browser, against 55.6% to 66.7% with Chrome DevTools CLI, Playwright CLI, Browser Use CLI and agent-browser. It also had the lowest average time, 324.6 seconds per task, and cost, $1.702, on the 13 tasks all five finished. The lead is widest on long jobs across many sites. Full setup and per-task results are in How we tested and Full results.

Which browser tool makes Claude Opus 5 fastest?

ego-browser had the lowest average time. On the 13 tasks all five tools finished, Claude Opus 5 averaged 324.6 seconds with ego-browser, 369.5 with Chrome DevTools CLI, 407.7 with Browser Use CLI, 495.8 with Playwright CLI and 543.6 with agent-browser. ego-browser needed the fewest tool calls, 30.3 per task against 41.2 to 53.7. Chrome DevTools CLI was faster on 5 of the 13 tasks. Full setup and per-task results are in How we tested and Full results.

Which browser tool is cheapest to run with Claude Opus 5?

ego-browser had the lowest average cost. On the 13 tasks all five tools finished, Claude Opus 5 cost $1.702 per task with ego-browser, $1.845 with Chrome DevTools CLI, $1.975 with Browser Use CLI, $2.052 with Playwright CLI and $2.368 with agent-browser. Over all 27 attempts, ego-browser and Playwright CLI were close, $3.280 and $3.330. The model and its price are the same, so the difference is usage, mostly cache-read tokens. Full setup and per-task results are in How we tested and Full results.

Is agent-browser or Playwright CLI better for Claude Code browser automation?

In our runs with Claude Opus 5, Playwright CLI finished more tasks: 18 of 27 against 15 for agent-browser. On the 13 tasks every tool finished, Playwright CLI averaged 495.8 seconds and $2.052 per task, agent-browser 543.6 seconds and $2.368. Playwright CLI's average time includes one 2,127-second run. agent-browser ran out of the 100-step budget on 11 tasks, Playwright CLI on 8, mostly long multi-site jobs. Full setup and per-task results are in How we tested and Full results.

Chrome DevTools CLI vs Playwright CLI: which works better with Claude Opus 5?

They tied on success rate, 18 of 27 tasks each, with Claude Opus 5 in Claude Code. On the 13 tasks every tool finished, Chrome DevTools CLI averaged 369.5 seconds and $1.845 per task, Playwright CLI 495.8 seconds and $2.052. Playwright CLI's average time includes one 2,127-second run on a single task. Over all 27 attempts, Playwright CLI averaged $3.330 and Chrome DevTools CLI $3.401. Full setup and per-task results are in How we tested and Full results.

How does Browser Use CLI compare with other browser tools for Claude Code?

Browser Use CLI ranked third on average time and cost and fourth on success rate in our runs. With Claude Opus 5, it finished 17 of 27 tasks, against 24 for ego-browser and 15 to 18 for the others. On the 13 tasks all five finished, it averaged 407.7 seconds and $1.975 per task, third on both. It matched ego-browser's pass rate on cross-site lookups and had the lowest time on the shared research tasks, but ran out of steps on 7 long multi-site tasks. Full setup and per-task results are in How we tested and Full results.

Why do browser agents fail on long multi-site tasks?

In our runs, mostly by running out of steps, not by giving wrong answers. Claude Opus 5 had a 100-step budget per attempt across 11 tasks spanning 8 to 16 sites each. With Chrome DevTools CLI, Playwright CLI, Browser Use CLI and agent-browser it used it all before answering on 7 to 9 of the 11. With ego-browser that happened once, and the same work took fewer tool calls, 73.4 per attempt against 91.5 to 97.0. Full setup and per-task results are in How we tested and Full results.

How much does a Claude Opus 5 browser agent task cost?

About $1.70 to $2.37 per task on average on the 13 tasks every tool finished, and more on long multi-site jobs. On 13 tasks that every tool finished, Claude Opus 5 in Claude Code averaged $1.702 per task with ego-browser and up to $2.368 with agent-browser. Tasks spanning 8 to 16 sites averaged $4.940 to $5.531 per attempt across the five tools. The browser tool moves the cache-read part of the bill most. Full setup and per-task results are in How we tested and Full results.

How long does a Claude Opus 5 browser agent take per task?

About 325 to 545 seconds on average for tasks every tool finished, and longer on multi-site jobs. On those 13 tasks, Claude Opus 5 averaged 324.6 seconds per task with ego-browser and 543.6 with agent-browser. Tasks across 8 to 16 sites averaged 919.1 to 1,112.2 seconds per attempt. Across the five tools, tool time varied more than model time. Full setup and per-task results are in How we tested and Full results.

Does the browser tool change Claude Opus 5's success rate?

Yes, on long tasks. With the same model and the same Claude Code setup, Claude Opus 5's success ranged from 55.6% with agent-browser to 88.9% with ego-browser across 27 tasks. On business web apps all five passed every task, and on cross-site lookups three of the five passed every task. The spread came from long multi-site jobs, where with four of the tools Opus 5 used all 100 steps before answering on 7 to 9 of 11 tasks. Full setup and per-task results are in How we tested and Full results.

How we tested

Setup

We held the model and agent fixed and changed the browser tool. Every run used Claude Opus 5 (claude-opus-5) at high reasoning effort in Claude Code, with the Bash, Read, Write, Edit and Skill tools and a skill file for each browser tool. ego-browser drives ego(lite) through scripts the agent writes. Chrome DevTools CLI, Playwright CLI, Browser Use CLI and agent-browser each drive Google Chrome through their own commands. Each attempt had a budget of 100 steps, where a step is one agent turn, and 60 minutes at most. Every turn before the final answer made one tool call. Each tool ran all 27 tasks.

Tasks

27 tasks on real websites and test web apps, in four types of work. Multi-site data collection has 11 tasks that gather the same fields from 8 to 16 official sites, such as the first 2026 rate decision at 12 central banks. Cross-site lookups has 9 tasks across one to three sites, such as checking FRED inflation series against BLS. Open-ended research has 3 tasks, such as reading points off a chart by hovering. Business web apps has 4 tasks in test apps, such as booking a trip within a travel policy. A 28th task, a SaaS compliance matrix, was dropped for every tool because one vendor changed its trust page partway through testing.

What counts as a pass

Each task has scored items, 190 in all. 161 have values fixed in advance, 18 come from data seeded in the test apps and 11 are checked against the live pages the agent saw. An attempt passes only if its final answer gets every scored item right. An attempt that ends without a final answer fails.

How we measured

An attempt is one run of one task with one tool. Each task ran once per tool, so the success rate counts attempts, and time and cost per task are those of that attempt. Speed and cost are compared on the 13 tasks all five tools passed. Time is wall time per attempt, with time waiting on the model and time inside the browser tool timed separately. Cost is computed from the tokens Claude Code reported, at Anthropic's list prices for Opus 5. The model is the same in every tool, so cost differences come from usage alone.

Full results

Update log: first published with results for all 27 tasks in each of the five browser tools.