# Testing Claude in Chrome, Claude Desktop and the Codex app’s browsers on the same job

Published 8 Nov 2026 · 10 min read

I ran the browser job from my agent-browser, Playwright and Chrome DevTools study with the four browsers that come with Claude’s and Codex’s own apps, and compared every run with that study’s.

My [study of agent-browser, Playwright and Chrome DevTools](https://maximilianfabian.dev/blog/browser-tools-for-agents/) left out the browsers that come with Claude’s and Codex’s own apps, because building production workflows on them ties you to one vendor’s agents and models. I still think that, but I wanted to know how they do on the same job. So I ran four of them: Claude in Chrome with Claude Code, Claude Desktop’s built-in browser, and the Codex app’s Chrome extension and built-in browser.

[The job](https://maximilianfabian.dev/blog/browser-tools-for-agents/#the-job) is the first study’s, with the script it gives the agent to read the report’s rows. An agent picks an option from a menu in a signed-in web app, opens a report in a new tab, saves the report’s hundreds of rows to a file, closes the tab and goes back to the start page. Each setup has 12 runs in the comparison. I compare each one with the first study’s nine setups that ran the same model: its three tools, each as a CLI with its skill, a CLI without it and an MCP server. Claude’s two browsers ran Opus 5.5, and the Codex app’s two ran GPT 6 Astra.

## Key results

1. With Opus 5.5, Claude in Chrome was slower and cost more than all three of the first study’s tools.
2. Claude Desktop’s built-in browser was faster and cost less than Claude in Chrome, but was slower than those three tools, and nearly always cost more.
3. With GPT 6 Astra, the Codex app’s browsers took about as long as those three tools, and mostly cost more.

After these runs, I’d still build production workflows on agent-browser, Playwright or Chrome DevTools.

**Every run passed except one with Claude in Chrome**

Runs that passed every check, out of 12 per setup, in the order they ran on 7 and 8 November 2026. The Codex app’s runs had the fixed script line.

Grid of squares, one per run in run order, filled for a pass. Every run passed except one with Claude in Chrome. Each setup’s count follows in a list.

- Claude Code with Claude in Chrome: 11 of 12 passed, run C2 failed.
- Claude Desktop with its built-in browser: 12 of 12 passed.
- Codex app with its Chrome extension: 12 of 12 passed.
- Codex app with its built-in browser: 12 of 12 passed.

Source: [my study runs](https://maximilianfabian.dev/blog/built-in-browsers-for-agents/#every-run), 7 and 8 November 2026.

## Where the time went

**Claude’s two browsers took longer than the first study’s tools**

Each run’s time in seconds. Both model groups share one scale. A first-study row is one tool’s 36 runs with that model, over its three connections. Every other row is one setup’s 12 runs. Lower is better.

Strip chart, one mark per run. Claude’s two browsers took longer than the first study’s tools. Each row’s range, median and passes follow in a list.

Axis: 0s, 100s, 200s, 300s, 400s, 500s

- Opus 5.5, medium

  - Claude Code with agent-browser in the first study: 36 runs, 73 seconds to 170 seconds, median 102 seconds, 36 of 36 passed.
  - Claude Code with Playwright in the first study: 36 runs, 74 seconds to 128 seconds, median 96 seconds, 36 of 36 passed.
  - Claude Code with Chrome DevTools in the first study: 36 runs, 62 seconds to 120 seconds, median 85 seconds, 36 of 36 passed.
  - Claude Code with Claude in Chrome: 12 runs, 195 seconds to 395 seconds, median 346 seconds, 11 of 12 passed.
  - Claude Desktop with its built-in browser: 12 runs, 119 seconds to 281 seconds, median 138 seconds, 12 of 12 passed.
- GPT 6 Astra, medium

  - Codex with agent-browser in the first study: 36 runs, 59 seconds to 127 seconds, median 79 seconds, 36 of 36 passed.
  - Codex with Playwright in the first study: 36 runs, 66 seconds to 160 seconds, median 93 seconds, 36 of 36 passed.
  - Codex with Chrome DevTools in the first study: 36 runs, 60 seconds to 277 seconds, median 80 seconds, 36 of 36 passed.
  - Codex app with its Chrome extension: 12 runs, 67 seconds to 103 seconds, median 82 seconds, 12 of 12 passed.
  - Codex app with its built-in browser: 12 runs, 79 seconds to 109 seconds, median 88 seconds, 12 of 12 passed.

Source: [my study runs](https://maximilianfabian.dev/blog/built-in-browsers-for-agents/#every-run), 7 and 8 November 2026, and [the first study’s runs](https://maximilianfabian.dev/blog/browser-tools-for-agents/#every-run), 19 to 22 October 2026.

Most of Claude in Chrome’s extra time went on getting the report’s rows from the page into the file. None of its tools saves text to a file. In every run that passed, the agent read the rows into its context and typed them into the file as model output. That took a median 199s.

Those setups got the rows to the file another way. Their agents redirected a CLI’s output to a file, used a tool’s option to save a result, or took the rows from a file that Claude Code had saved. Claude Code saves a tool result that is over its token limit to a file and gives the agent the path.

Claude Desktop’s browser took that last route. Its script tool returned all the rows in one result, and Claude Code saved it to a file. In 9 of 12 runs the agent copied the rows from that file with a script. In the other 3, the agent first shrank the rows to about 25,000 characters, which came back whole, and typed them out. Two of those agents said they expected a full result to be cut or to be too large. Those runs took a median 271s and cost $1.62, against 130s and $0.80 for the other nine.

The rows are not the whole gap. Tool calls took a median 62s of a run with Claude in Chrome and 44s with Claude Desktop’s browser, against 4s to 18s for the first study’s Claude Code setups. The run records don’t show what inside a call took the time.

In the Codex app, the agent’s JavaScript runs in Node.js and drives the page through Playwright. So one call ran the job’s script in the page and wrote its result to the file. A median run made 8 browser commands, against 14.5 to 27 for the first study’s Codex setups.

## Where the money went

**Claude’s two browsers cost more than the first study’s tools**

Each run’s cost in US dollars. Both model groups share one scale. A first-study row is one tool’s 36 runs with that model, over its three connections. Every other row is one setup’s 12 runs. Claude Code and Claude Desktop report their own cost, and I priced Codex’s tokens at list prices. Lower is better.

Strip chart, one mark per run. Claude’s two browsers cost more than the first study’s tools. Each row’s range, median and passes follow in a list.

Axis: $0.00, $1.00, $2.00, $3.00

- Opus 5.5, medium

  - Claude Code with agent-browser in the first study: 36 runs, 0.26 dollars to 1.13 dollars, median 0.48 dollars, 36 of 36 passed.
  - Claude Code with Playwright in the first study: 36 runs, 0.27 dollars to 0.52 dollars, median 0.34 dollars, 36 of 36 passed.
  - Claude Code with Chrome DevTools in the first study: 36 runs, 0.27 dollars to 0.46 dollars, median 0.33 dollars, 36 of 36 passed.
  - Claude Code with Claude in Chrome: 12 runs, 0.58 dollars to 1.62 dollars, median 1.38 dollars, 11 of 12 passed.
  - Claude Desktop with its built-in browser: 12 runs, 0.76 dollars to 1.63 dollars, median 0.82 dollars, 12 of 12 passed.
- GPT 6 Astra, medium

  - Codex with agent-browser in the first study: 36 runs, 0.54 dollars to 1.00 dollars, median 0.86 dollars, 36 of 36 passed.
  - Codex with Playwright in the first study: 36 runs, 0.55 dollars to 0.92 dollars, median 0.66 dollars, 36 of 36 passed.
  - Codex with Chrome DevTools in the first study: 36 runs, 0.46 dollars to 0.89 dollars, median 0.57 dollars, 36 of 36 passed.
  - Codex app with its Chrome extension: 12 runs, 0.84 dollars to 0.98 dollars, median 0.85 dollars, 12 of 12 passed.
  - Codex app with its built-in browser: 12 runs, 0.80 dollars to 1.29 dollars, median 0.92 dollars, 12 of 12 passed.

Source: [my study runs](https://maximilianfabian.dev/blog/built-in-browsers-for-agents/#every-run), 7 and 8 November 2026, and [the first study’s runs](https://maximilianfabian.dev/blog/browser-tools-for-agents/#every-run), 19 to 22 October 2026.

Claude in Chrome cost the most because of the rows. Its median run produced 34.1K output tokens, against 12.3K with Claude Desktop’s browser and 5.7K to 8.7K for the first study’s Claude Code setups.

Claude Desktop’s browser still cost more than eight of the first study’s nine Claude Code setups, partly because every model call was large. A call sent a median 80K input tokens, against 16K to 73K for those setups. Claude Desktop also keeps its cache for an hour, which costs more per token than Claude Code’s five-minute cache.

The Codex app’s two browsers cost more than seven of the first study’s nine Codex setups, partly because their model calls were large too. A call sent a median 37K and 38K input tokens, against 17K to 32K for those setups. The first call held about 23K tokens, and later calls grew when the agent read the page’s state.

## What failed

One line of the job’s script stopped the Codex app’s browsers. The line read `getComputedStyle` through `globalThis`, which the Codex app’s page scope doesn’t have, so it threw an error. In 20 of the first 24 Codex runs the agent stopped there, as the job tells it to, and the run failed. I fixed the line, ran both Codex setups 12 more times and use only those runs here. The fixed line gives the same result wherever `globalThis` exists, as it does in Claude’s browsers and the first study’s tools.

Claude in Chrome’s one failed run lost its rows at Claude Code’s auto-mode classifier. The agent packed the rows into a compressed format of its own and asked for it in nine pieces. The classifier refused that call and gave “Auto-Mode Bypass” as its reason, although Claude Code ran with its permission checks skipped. The agent did not retry, so it saved no rows. It finished the rest of the job and recorded the refusal. The classifier refused no call in any other run. All 216 of the first study’s runs passed.

## Limits

I measured one job in one web app, at medium effort, with 12 runs per setup, on one machine. Another job could order the setups differently. This one has the agent save hundreds of rows, which favours a tool that can send a result to a file.

The two studies also differ in these ways:

- For the three desktop-app setups, a run’s time started when I sent the instructions. For Claude in Chrome and the first study, it started when Claude Code or Codex launched. That favours the desktop apps.
- With Claude in Chrome and Claude Desktop’s browser, the agent had to open the start page itself. That extra step counts against those two.
- Three more differences could push either way. Claude in Chrome and the Codex app’s Chrome extension drove a visible Chrome window, where the first study ran Chrome headless. The desktop apps ran newer versions of Claude Code and Codex. They also ran with their own permission settings, where Claude Code skipped its permission checks.
- The first study pooled three operating systems, and these runs used one machine. I state only differences that also hold against each operating system’s runs, which is why some sentences compare with its three tools and others with its nine setups.

## How I ran the study

**Show the method**

I used these versions:

- Claude Code 2.1.303 with `claude-opus-5-5`, medium effort, and the Claude in Chrome extension 1.0.105
- Claude Desktop 4.16336.0, which runs Claude Code 2.1.317, with `claude-opus-5-5`, medium effort
- The ChatGPT desktop app 26.1029.61003, which runs Codex 0.166.0-alpha.6, with `gpt-6-astra`, medium effort, and its Chrome extension 1.26.1102.11740
- Google Chrome 155, stable release, with one study profile

The first study ran Claude Code 2.1.303 and Codex 0.163.0-alpha.11.

The ChatGPT desktop app runs Codex, so this post calls it the Codex app. Claude Desktop labels its built-in browser the Browser pane. A desktop app here means Claude Desktop or the Codex app, and the web app is the one the job works in. Claude in Chrome and the Codex app’s Chrome extension drove the study’s Chrome. The built-in browsers in Claude Desktop and the Codex app are the desktop apps’ own browsers.

Claude Code ran with the first study’s command line, which skips permission checks and loads no other MCP servers or settings, plus the flag that connects Claude in Chrome. The other three setups ran with their desktop app’s own settings. Claude Desktop ran in its Auto permission mode, and the Codex app with approval on request, its automatic reviewer and a sandbox limited to the run’s folder. Each run was a fresh session or thread in a folder that held only the job’s files, and neither desktop app kept memories between runs.

Every setup got the first study’s instructions. I rewrote only the part that names the browser, and for two of the browsers I added a step that opens the start page, because the job starts with a signed-in tab on that page:

- Claude in Chrome acts only on tabs in its own tab group, so the agent opened the start page in a new tab in its group.
- Claude Desktop gives each session its own built-in browser, which has no tabs until the agent first opens a page. The agent opened the start page there. The browser also blocks tabs that a page opens by itself, so the agent opened the report’s link in a new tab instead.
- The Codex app’s Chrome extension took over the tab that was already open in Chrome.
- The Codex app’s built-in browser opened on the start page when the thread started.

The pass checks were the same as in the first study, which also checked that the report’s tab was closed. A run passed only if it saved rows that matched a reference exactly, left the web app in the right state, closed the report’s tab, used only its assigned browser, never asked me for input and finished within 30 minutes. In one Claude in Chrome run, the study’s Chrome opened a `chrome://inspect` tab by itself, and no agent call asked for it. I counted that tab as the browser’s own, so the run passed.

Before the job’s script read a table, it checked that the table was visible:

```js
const style = globalThis.getComputedStyle ? getComputedStyle(element) : null;
```

The Codex app runs a page’s JavaScript in a scope that has `getComputedStyle`, `window` and `document` but no `globalThis`. In the 4 first Codex runs that passed, the agent checked what the scope had and wrapped the script in a function that defined `globalThis`. I replaced the check with `typeof getComputedStyle === 'function'`.

Two other Claude in Chrome runs also compressed the rows before reading them out, and the classifier refused neither.

I signed each browser into the web app before the runs. All runs took place on 7 and 8 November 2026, one at a time. The Claude in Chrome runs and the first Codex runs went in rounds of one run per setup, with the order rotating from round to round. The Codex app’s runs with the fixed line came after them, alternating which of the two went first. Claude Desktop’s 12 runs came last, on 8 November, one after another. The first study’s runs took place from 19 to 22 October 2026.

Each number is a setup’s median over its passed runs unless I say otherwise. I took each measure’s median on its own, so a median run can mix values from different runs. A cache share is the exception, counting all of a setup’s passed runs together. The charts plot every run.

A run’s tool time is the time inside its tool calls, and its model time is the rest of the run. A browser command is one call to a browser tool. In the Codex app, one call runs a JavaScript program that can take several steps on the page, and Claude in Chrome has a tool that runs several steps in one call, so their counts run lower. Token counts come from the logs of Claude Code, Claude Desktop and the Codex app. Input tokens include cache reads, and output tokens include reasoning.

What this post says about how a browser’s tools behave comes from the tool results and agent messages in the run records, and from the tool definitions: Claude Desktop’s and the Codex app’s as their records hold them, and Claude in Chrome’s as the Claude Code 2.1.303 it ran with holds them. These figures come from the same records:

- With Claude Desktop’s browser, the script’s result came back whole at about 70,000 characters, which Claude Code saved to a file.
- How the first study’s agents saved the rows comes from 36 of its 108 Claude Code runs. None of the 108 produced more than 10.6K output tokens, fewer than any Claude in Chrome run that passed.

Claude Code and Claude Desktop report what each run would cost through the API, and I priced Codex’s tokens at list prices. None of these is what you’d pay on a subscription plan. A median run with Claude Desktop’s browser wrote 38K tokens to its cache and read 1.38M from it, against 91K and 1.22M with Claude in Chrome. Cache reads made up 96.8% of input tokens with Claude Desktop’s browser, 93.8% with Claude in Chrome, 93.2% with the Codex app’s Chrome extension and 92.2% with its built-in browser. A median run with the Codex app used 480K and 511K input tokens and produced 1.5K and 1.7K output tokens, more input and less output than any of the first study’s Codex setups.

The web app’s name and the report’s contents stay private, so this post reports only derived measurements.

## Every setup’s median run

**Claude in Chrome was the slowest and costliest setup**

The first study’s nine Claude Code setups, then Claude’s two browsers, all with Opus 5.5 at medium effort and 12 runs each. Each number is the median of a setup’s passed runs. Costs are what Claude Code and Claude Desktop report.

| Setup | Time | Input tokens | Cost | Output tokens | Model calls | Browser commands | Passed |
| --- | --- | --- | --- | --- | --- | --- | --- |
| agent-browser, first study |   |   |   |   |   |   |   |
| CLI + skill | 104s | 652K | $0.48 | 6.8K | 22 | 40 | 12 of 12 |
| CLI, no skill | 96s | 304K | $0.28 | 5.9K | 20 | 38 | 12 of 12 |
| MCP | 106s | 1.59M | $1.02 | 8.7K | 21 | 27 | 12 of 12 |
| Playwright, first study |   |   |   |   |   |   |   |
| CLI + skill | 90s | 332K | $0.32 | 5.8K | 16 | 22 | 12 of 12 |
| CLI, no skill | 99s | 353K | $0.33 | 6.6K | 20 | 25 | 12 of 12 |
| MCP | 98s | 513K | $0.41 | 7.3K | 21 | 15 | 12 of 12 |
| Chrome DevTools, first study |   |   |   |   |   |   |   |
| CLI + skill | 97s | 341K | $0.32 | 6.1K | 17 | 22 | 12 of 12 |
| CLI, no skill | 98s | 348K | $0.32 | 5.9K | 19 | 24 | 12 of 12 |
| MCP | 72s | 517K | $0.40 | 5.7K | 18 | 15 | 12 of 12 |
| This study |   |   |   |   |   |   |   |
| Claude Code with Claude in Chrome | 346s | 1.31M | $1.38 | 34.1K | 30 | 23 | 11 of 12 |
| Claude Desktop with its built-in browser | 138s | 1.42M | $0.82 | 12.3K | 18 | 28 | 12 of 12 |

Source: [my study runs](https://maximilianfabian.dev/blog/built-in-browsers-for-agents/#every-run), 7 and 8 November 2026, and [the first study’s runs](https://maximilianfabian.dev/blog/browser-tools-for-agents/#every-run), 19 to 22 October 2026.

**The Codex app’s browsers used the most input tokens**

The first study’s nine Codex setups, then the Codex app’s two browsers, all with GPT 6 Astra at medium effort and 12 runs each. Each number is the median of a setup’s passed runs. I priced Codex’s tokens at list prices.

| Setup | Time | Input tokens | Cost | Output tokens | Model calls | Browser commands | Passed |
| --- | --- | --- | --- | --- | --- | --- | --- |
| agent-browser, first study |   |   |   |   |   |   |   |
| CLI + skill | 91s | 426K | $0.91 | 2.2K | 14 | 27 | 12 of 12 |
| CLI, no skill | 78s | 289K | $0.68 | 1.9K | 14 | 23 | 12 of 12 |
| MCP | 72s | 414K | $0.89 | 1.8K | 13 | 21 | 12 of 12 |
| Playwright, first study |   |   |   |   |   |   |   |
| CLI + skill | 102s | 318K | $0.80 | 2.2K | 16 | 16 | 12 of 12 |
| CLI, no skill | 91s | 243K | $0.66 | 2.2K | 14 | 18 | 12 of 12 |
| MCP | 88s | 261K | $0.62 | 2.1K | 14 | 18 | 12 of 12 |
| Chrome DevTools, first study |   |   |   |   |   |   |   |
| CLI + skill | 75s | 197K | $0.53 | 1.9K | 11 | 17 | 12 of 12 |
| CLI, no skill | 86s | 233K | $0.58 | 1.9K | 13 | 18 | 12 of 12 |
| MCP | 81s | 238K | $0.60 | 2.0K | 12 | 15 | 12 of 12 |
| This study |   |   |   |   |   |   |   |
| Codex app with its Chrome extension | 82s | 480K | $0.85 | 1.5K | 13 | 8 | 12 of 12 |
| Codex app with its built-in browser | 88s | 511K | $0.92 | 1.7K | 13 | 8 | 12 of 12 |

Source: [my study runs](https://maximilianfabian.dev/blog/built-in-browsers-for-agents/#every-run), 7 and 8 November 2026, and [the first study’s runs](https://maximilianfabian.dev/blog/browser-tools-for-agents/#every-run), 19 to 22 October 2026.

## Every run

One row per run of this study, with time in seconds and cost in US dollars. Rows to file is the time from the tool result that reported the extracted rows to the end of the call that wrote them to the file. Claude’s runs are numbered C1 to C12 in the order they ran. The Codex app’s runs, F1 to F12, ran with the fixed line. The first study’s runs are in [its every-run table](https://maximilianfabian.dev/blog/browser-tools-for-agents/#every-run).

**Every run**

| Run | Result | Time (s) | Rows to file (s) | Input tokens | Cached input | Output tokens | Cost (USD) | Model calls | Browser commands |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Claude Code with Claude in Chrome |   |   |   |   |   |   |   |   |   |
| C1 | passed | 392.6 | 219.5 | 1,901,725 | 1,809,463 | 35,987 | $1.54 | 39 | 28 |
| C2 | failed (saved no matching rows) | 230.9 | not measured | 710,318 | 665,307 | 11,028 | $0.58 | 23 | 20 |
| C3 | passed | 341.5 | 198.7 | 1,101,985 | 1,009,341 | 34,147 | $1.35 | 26 | 22 |
| C4 | passed | 194.9 | 82.1 | 791,091 | 757,680 | 14,171 | $0.60 | 24 | 19 |
| C5 | passed | 375.0 | 219.3 | 1,926,670 | 1,819,369 | 35,889 | $1.62 | 33 | 26 |
| C6 | passed | 350.7 | 213.6 | 1,307,273 | 1,219,139 | 34,909 | $1.38 | 30 | 23 |
| C7 | passed | 361.7 | 229.0 | 1,608,969 | 1,510,524 | 37,186 | $1.54 | 33 | 23 |
| C8 | passed | 321.5 | 193.2 | 1,015,645 | 925,129 | 33,784 | $1.31 | 25 | 22 |
| C9 | passed | 345.7 | 183.9 | 1,278,793 | 1,210,974 | 32,676 | $1.24 | 30 | 24 |
| C10 | passed | 330.7 | 185.2 | 1,429,199 | 1,334,995 | 32,462 | $1.39 | 30 | 25 |
| C11 | passed | 312.2 | 188.7 | 1,291,330 | 1,204,867 | 30,035 | $1.27 | 28 | 25 |
| C12 | passed | 394.7 | 260.4 | 1,451,409 | 1,362,735 | 41,237 | $1.54 | 28 | 23 |
| Claude Desktop with its built-in browser |   |   |   |   |   |   |   |   |   |
| C1 | passed | 129.0 | 18.2 | 1,257,456 | 1,221,795 | 11,423 | $0.76 | 16 | 27 |
| C2 | passed | 256.9 | 136.5 | 1,607,155 | 1,528,768 | 31,525 | $1.56 | 18 | 27 |
| C3 | passed | 129.8 | 15.4 | 1,334,836 | 1,298,333 | 11,503 | $0.78 | 17 | 29 |
| C4 | passed | 271.1 | 157.7 | 1,855,273 | 1,778,281 | 32,551 | $1.62 | 20 | 23 |
| C5 | passed | 202.6 | 15.7 | 1,630,811 | 1,583,752 | 18,839 | $1.07 | 20 | 28 |
| C6 | passed | 138.4 | 13.0 | 1,415,181 | 1,378,071 | 12,164 | $0.82 | 18 | 28 |
| C7 | passed | 137.8 | 14.8 | 1,350,282 | 1,312,713 | 11,989 | $0.80 | 17 | 25 |
| C8 | passed | 118.5 | 14.0 | 1,274,727 | 1,236,205 | 11,671 | $0.79 | 16 | 27 |
| C9 | passed | 119.5 | 12.4 | 1,346,885 | 1,309,823 | 11,555 | $0.79 | 17 | 28 |
| C10 | passed | 280.8 | 162.7 | 1,807,240 | 1,730,759 | 33,813 | $1.63 | 20 | 29 |
| C11 | passed | 164.4 | 27.8 | 1,633,788 | 1,593,459 | 15,171 | $0.94 | 20 | 26 |
| C12 | passed | 129.6 | 14.4 | 1,422,366 | 1,384,842 | 12,418 | $0.83 | 18 | 28 |
| Codex app with its Chrome extension |   |   |   |   |   |   |   |   |   |
| F1 | passed | 75.2 | 0.0 | 479,354 | 438,144 | 1,473 | $0.92 | 13 | 8 |
| F2 | passed | 68.1 | 0.0 | 480,000 | 448,512 | 1,513 | $0.84 | 13 | 8 |
| F3 | passed | 89.2 | 0.0 | 573,117 | 539,520 | 2,026 | $0.98 | 15 | 10 |
| F4 | passed | 74.2 | 0.0 | 480,508 | 448,768 | 1,492 | $0.84 | 13 | 8 |
| F5 | passed | 67.1 | 0.0 | 483,392 | 450,560 | 1,379 | $0.85 | 13 | 8 |
| F6 | passed | 103.4 | 0.0 | 481,014 | 449,408 | 1,873 | $0.86 | 13 | 8 |
| F7 | passed | 100.2 | 0.0 | 572,100 | 538,624 | 1,644 | $0.96 | 15 | 10 |
| F8 | passed | 88.3 | 0.0 | 480,511 | 448,000 | 1,537 | $0.85 | 13 | 8 |
| F9 | passed | 82.0 | 0.0 | 478,811 | 447,232 | 1,741 | $0.85 | 13 | 8 |
| F10 | passed | 81.0 | 0.0 | 479,979 | 436,992 | 1,682 | $0.95 | 13 | 8 |
| F11 | passed | 83.3 | 0.0 | 479,961 | 448,512 | 1,546 | $0.84 | 13 | 8 |
| F12 | passed | 76.2 | 0.0 | 480,087 | 448,512 | 1,478 | $0.84 | 13 | 8 |
| Codex app with its built-in browser |   |   |   |   |   |   |   |   |   |
| F1 | passed | 91.4 | 0.0 | 487,650 | 456,576 | 1,550 | $0.84 | 13 | 8 |
| F2 | passed | 78.6 | 0.0 | 533,783 | 492,416 | 1,553 | $0.98 | 14 | 9 |
| F3 | passed | 86.6 | 0.0 | 445,005 | 414,592 | 1,657 | $0.80 | 12 | 7 |
| F4 | passed | 81.1 | 0.0 | 533,952 | 492,672 | 1,721 | $0.99 | 14 | 9 |
| F5 | passed | 95.9 | 0.0 | 490,785 | 459,392 | 1,655 | $0.86 | 13 | 8 |
| F6 | passed | 109.0 | 0.0 | 530,558 | 489,344 | 1,662 | $0.98 | 14 | 9 |
| F7 | passed | 96.9 | 0.0 | 491,451 | 460,288 | 1,706 | $0.86 | 13 | 8 |
| F8 | passed | 101.3 | 0.0 | 490,919 | 448,000 | 1,560 | $0.96 | 13 | 8 |
| F9 | passed | 81.3 | 0.0 | 530,499 | 499,200 | 1,599 | $0.89 | 14 | 9 |
| F10 | passed | 89.0 | 0.0 | 615,657 | 549,760 | 1,701 | $1.29 | 13 | 8 |
| F11 | passed | 83.4 | 0.0 | 531,202 | 469,888 | 1,718 | $1.17 | 14 | 9 |
| F12 | passed | 87.2 | 0.0 | 491,820 | 460,544 | 1,762 | $0.86 | 13 | 8 |

Source: [my study runs](https://maximilianfabian.dev/blog/built-in-browsers-for-agents/#every-run), 7 and 8 November 2026.

## Cite this study

Maximilian Fabian Kirchner (2026). Testing Claude in Chrome, Claude Desktop and the Codex app’s browsers on the same job. maximilianfabian.dev, published 8 November 2026. https://maximilianfabian.dev/blog/built-in-browsers-for-agents/

```bibtex
@misc{kirchner2026-built-in-browsers-for-agents,
  author       = {Kirchner, Maximilian Fabian},
  title        = {{Testing Claude in Chrome, Claude Desktop and the Codex app’s browsers on the same job}},
  howpublished = {\url{https://maximilianfabian.dev/blog/built-in-browsers-for-agents/}},
  year         = {2026},
  month        = nov,
  note         = {Published 8 November 2026},
}
```
