An outside study of Perplexity Computer by Abhinav Gupta
27 runs of Perplexity Computer on my own account, Oct 2 to Oct 8, 2026. Each chart backs one of the three problems. Red is what was wrong or charged later, green is what matched.
Counts from my runs, not rates.
Price comparison, 3 apps, Mac app, cloud. The first part of each bar is the cheapest run; red is what the others cost on top.
3 runs of one prompt. It shows the price can move, not how often.
Effort set to High, against Standard, on the web. 3.9x the credits.
One run on each setting, web only.
The panel does catch up later.
One column per run, lowest first. Blue is what the panel showed when the run said Finished. Red is the part of the charge that landed later.
Shown at FinishedCharged laterAlready matched
Readings taken at the Finished screenshot. Final charges read from Analytics afterwards.
The same session, on the Mac app and then on the web about half a minute later.
One Mac, one local model (Gemma 4 E4B).
One dot per run with Hybrid on. Filled means the local model ran. It ran only when Perplexity’s privacy prompt (the Privacy Gate) appeared, and that prompt flagged the bank file and nothing else.
In the other runs nothing ran locally. One of them, lab results flags from a csv, cost 8.23 credits with Hybrid on against 3.92 in the cloud.
Bank report, Hybrid run 2. Each bar is what the report said, as a share of the real total. The line marks the real total.
What the Hybrid report saidLeft out of the reportThe real total
The cloud run of the same task got 8 of 8. In 3 of the 3 runs where the local model ran, the result was wrong or incomplete. On the Excel budget, 0 of 40 rows came back complete.
Excel budget, Hybrid run 22. The values it left blank were on the Mac the whole time.
| Task | Privacy Gate | Ran on the Mac | Cloud credits | Hybrid credits | Cloud time | Hybrid time | Cloud result | Hybrid result |
|---|---|---|---|---|---|---|---|---|
| Bank report from a CSV | appeared | yes | 15.02 to 15.55 | 11.77 to 11.87 | 30 s to 60 s* | 7.0 min to 9.5 min* | 6 of 6; 8 of 8 | 1 of 6; 0 of 8 |
| Excel budget from a CSV | appeared | yes | 12.63 to 14.60 | 15.91 | 32 s (1 of 2 timed) | 6.3 min* | 40 of 40 rows | 37 of 40 amounts, 0 of 40 dates |
| Invoice summary from a CSV | did not appear | no | 11.29 | 10.94 | 28 s | 27 s | 7 of 7 totals and due dates | 7 of 7 totals and due dates |
| Lab results flags from a CSV | did not appear | no | 3.92 | 8.23 | 10 s | 19 s | 8 of 8 | 8 of 8 |
| One-page website | no file | no | 16.45 | 15.63 | 43 s | 41 s | 4 of 4 parts | 4 of 4 parts |
| 5-slide deck | no file | no | 18.89 to 20.24 | 20.67 | 37 s to 41 s (2 of 3 timed) | 45 s | 5 of 5 slides | 5 of 5 slides |
| Comparison of 5 apps | no file | no | 32.25 | 42.06 | not timed | 42 s | not graded | not graded |
* Includes times read from screenshot timestamps, not a stopwatch.
500 recent App Store reviews, Reddit threads and press coverage. Later, 14,791 store reviews in a dashboard.
6 problem statements. The largest group was about credits. I ruled out exporters, browser clones and UI redesigns.
I walked through Computer, Hybrid and Skills on my own account.
27 runs with made-up files and answer keys, Oct 2 to Oct 8, 2026.
Run Receipt first, then Session to Skill.
Every file is made up: a bank export of 40 transactions, an invoice list and a set of lab results. Each has an answer key, kept out of the test folder so it could not be attached by mistake.
For each run I recorded the panel reading when it said Finished, the final charge from Analytics, the time, the files it produced, and whether the Privacy Gate appeared. I graded each output against its key afterwards.
“Hybrid” means the Hybrid toggle was on, with Gemma 4 E4B as the local model. “Cloud” means it was off. 24 planned tasks and 3 early tests, 694.46 credits ($6.94) in total. Task types: Bank report from a CSV; Excel budget from a CSV; Invoice summary from a CSV; Lab results flags from a CSV; Three short emails; Web pricing summary; Price comparison, 3 apps; One-page website; 5-slide deck; Two-page research brief; Comparison of 5 apps.
Small numbers. Most comparisons are one run on each side, on one account, one Mac and one local model, in one week. Nothing here is a rate.