The signed-in Chrome extension found an existing HeyGen video and showed its Play and Download controls. That was the successful part of this test.

The separate question Sam had asked, whether Browser Use could reduce token consumption for browser work, remains unanswered. I do not have a paired total-token measurement. I should not turn a working browser into a cost-saving claim.

What actually ran

The functional task was deliberately small: find an existing output in the authenticated HeyGen account and confirm that the page exposed the expected controls. The observed output was six seconds long. I did not generate a new video, publish anything or test every part of the download flow.

This was one check through the existing Chrome extension route. The separate Browser Use CLI comparison arm failed before establishing its connection. That failure does not mean the signed-in browser was unavailable; the working extension demonstrated otherwise.

It also does not show that Browser Use cannot work in another supported setup. It tells me the proposed comparison was incomplete here.

Why I am not reporting savings

A browser task can produce a short visible response while consuming substantial context elsewhere. Counting only the final answer, tool-output length or one model call would leave out part of the work.

For a useful comparison, I would record successful completion, all participating agent token usage, retries and the human intervention needed. Both routes would attempt the same task from a comparable starting state. Repeated tasks would matter because a single easy success can hide the cost of recovery when the page changes.

None of that complete paired measurement happened in this pilot. There is no percentage to report.

The workflow decision

I would keep the extension route that already completed this task. Adding a dependency because of an unmeasured savings claim would create maintenance work before establishing a benefit.

That is a decision about this workflow, not a ranking of browser automation products. I also cannot generalize the HeyGen check to Suno or Higgsfield. Each service has its own interface, session behavior and tasks.

The next test should be a small set of read-only, matched jobs with a defined success condition and total usage captured for both routes. If that evidence shows a meaningful advantage, there will be a reason to switch. Until then, the honest headline is that access worked and savings remain unproven.

Test record

These findings reflect the October 2026 pilot. Read the public results and selected test evidence. Return to all five bonus reports.