Estimate hit rate
Runs whose final charge landed inside the range shown before sending ÷ runs shown a range. Weekly.
In my runs: 2 of 12, or 11 of 12 with the range 20% wider. Bar: 80%.
Prototype · independent, not made by or with Perplexity
Run Receipt
A receipt for every Computer run: the cost before you send, an honest count while it runs, numbers checked by code after.
By Abhinav Gupta, product manager (six years), MS Product Management at Carnegie Mellon. 27 runs on my own account, Oct 2 to Oct 8, 2026.
Problem
Same prompt, three runs: 12.14, 14.29 and 21.96 credits. Nothing on screen says which you'll get.
- 12.14 to 21.96credits for one prompt, same setting
- 23 of 25times the meter read low at Finished
- 3 of 3Hybrid runs on my Mac came back wrong or incomplete
202 paying users ran out of creditsbecause you don’t get an estimate it’s hard for budgeting.
2 of 5 stars, App Store review, app v26.31.0
Solution
One receipt, at the three moments that go wrong
Categorize my September bank export, total each category, flag duplicates and refunds, price-check my top 3 subscriptions, one-page report.
Before you send
- This run will cost
- 11.77 to 11.87 credits
- from 2 past runs of this task
- It will take
- 7.0 min to 9.5 min
Where it runsYour Mac · 3Cloud · 1Search · 1
Why a range: one prompt of mine cost 12.14 to 21.96 credits on the same setting.
Where each step will run
- 1Read your bank fileYour Mac
private file stays on the device
- 2Categorize 40 transactionsYour Mac
private rows, local model. Check: Every row gets one category.
- 3Add up totalsYour Mac
arithmetic on private rows. Check: Totals are checked against the file.
- 4Look up 3 subscription pricesSearch
needs live prices
- 5Write the reportCloud
no private values needed
The plan is the prototype's own. Perplexity shows one total per run, so the range is for the whole run, not per step.
Metrics
Runs land inside the quoted range, and people still send them
- Runs sent, against the week before≤5% drop
- Range width, median<50%
- Meter gap at Finished<2%
- Repair taken when a check failsafter baseline
The bars are my proposals. I don't have Perplexity's data, so each resets after a baseline week.
Trade-offs
What I gave up, on purpose
A range, not a price
Overone number people can hold you to
Costwide ranges help less (49% median in my runs)
Say “provisional” out loud
Overhiding the count until it settles
Costpeople still see a number that will change
Code checks the sums, not the model
Overasking the model to grade itself
Costonly works when there's a file to check against
Research
For anyone who wants to dig in
Three problems, measuredA chart for each, from my 27 runs
Cost is a surprise.
12.14 to 21.96credits for one prompt
The same prompt on the same setting cost 12.14 to 21.96 credits across 3 runs. Nothing on my screen said which price I would get.
Limit. 3 runs of one prompt. A count, not a rate.
The credit meter reads low while you watch it.
23 of 25readings below the final charge
When the run said Finished, the meter was below the final charge in 23 of 25 readings.
Limit. The meter does catch up later.
Hybrid mode is unpredictable.
3 of 8Hybrid runs ran on the Mac
Hybrid is meant to run private steps on your Mac. With it on, the local model ran in 3 of 8 runs. In 3 of those 3, the result was wrong or incomplete.
Limit. One Mac, one local model.
What users say14,791 store reviews, labelled
- 202reviews are from paying users who ran out of credits
- 159of those came with 1 or 2 stars
- 68reviews say Computer burned through their credits
This is not the largest complaint. Free question limits and upgrade prompts (507) and Wrong or invented answers (444) are bigger. Run Receipt is aimed at paying Computer users.
“because you don’t get an estimate it’s hard for budgeting.”
“It uses up credits fast with no way to stop prompts after they start”
From 14,791 Google Play and App Store reviews, Jul 11, 2026 to Oct 8, 2026. 5,115 say something specific. 1,059 are complaints and requests about credits, limits, pricing or billing. Public reviews, not interviews. Labels come from an LLM, checked against 120 hand-labelled reviews: 117 matched. A keyword search of these reviews found none about the credit meter reading low. That finding comes from my own runs.
Full metric planFormulas, baselines and the decision each one controls
| Metric | How it is counted | In my runs | Starting bar | The decision it controls |
|---|---|---|---|---|
| Estimate hit rate | Runs whose final charge landed inside the range shown before sending, divided by runs that showed a range. Weekly, tasks with past runs. | 2 of 12A range built from the other runs of the same prompt held the next run. Widened by 20% at each end: 11 of 12. | 80%of runs inside the range. | Below the bar, the range cannot be trusted yet. Widen it or hold the rollout. |
| Range width | High end minus low end, divided by the low end. Median across tasks, weekly. It keeps the hit rate honest: a very wide range always hits. | 49%Median with the 20% widening. Without it, 7%. The widest single prompt was 81%. | under 50%median width. | Above the bar, the range is too wide to budget with. Collect more past runs before showing it. |
| Meter gap at Finished | Final charge minus the credits shown at Finished, divided by the final charge. Median per run, weekly. | 18%Median of 25 readings. 23 of them read low. | under 2%median gap. | The provisional mark stays until the gap is under the bar. |
| Repair take-up | Runs where the user took the offered repair, divided by runs where a check failed. Weekly, file tasks. | No baselineOnly I have used the prototype. In my runs the local model ran 3 times and 3 came back wrong or incomplete. | Set after a baseline week | If few take it, the repair is priced or placed wrong. |
| Guardrail: runs sent | Runs sent, divided by prompts where a cost range was shown. Weekly, against the week before. | No baselineThe prototype replays recorded runs, so nobody chose whether to send. | 5%lower at most. | A bigger drop means the range is stopping runs people would have valued. Change how it is shown. |
What it could be worthPer run, and the levers only Perplexity can size
Per run, the money is small. At $0.01 a credit:
- $0.10between the cheapest and the most expensive run of one prompt (12.14 to 21.96 credits)
- $0.03median amount the meter read low at Finished (2.64 credits)
- $2.21for one prompt on High effort, against $0.57 on Standard
So the case is not cents saved per run. It is whether paying users who cannot predict their credits stay subscribed.
Lever 1
Fewer cancellations
Paying users who run out of credits without warning. Perplexity would plug in: paying Computer users, the share who hit a credit surprise, and how many of those cancel.
Lever 2
Fewer billing contacts
Refund and credit disputes that start with a charge higher than the screen showed. Perplexity would plug in: contacts tagged credits or billing, and the cost of one.
Lever 3
More top-ups, or fewer
A user who can budget may buy more credits. A user who sees the price may send fewer runs. The guardrail metric above is there to tell which.
Break-even. One retained subscriber-month is worth $20 on Pro and $200 on Max. Every $10,000 of build and running cost is paid back by 500 retained Pro months or 50 retained Max months.
I cannot size this. Subscriber counts, churn and credit revenue are inside Perplexity, and I will not guess them. The public signal is a count, not a rate: 202 store reviews in the window above are from paying users who ran out of credits, 159 of them at 1 or 2 stars. Plan prices are list prices from third-party pricing guides, not Perplexity's own plans page, checked Oct 9, 2026.
What this can't tell youLimits of a one-person study
- Small numbers. Most comparisons are one run on each side. Nothing here is a rate.
- One account, one Mac, one local model (Gemma 4 E4B), one week.
- These are my own runs. I have not interviewed Perplexity users, so how much these problems matter to them is still a hypothesis.
- The prototype is standalone. It replays recorded runs and does not hook into Perplexity's product.
Every table, chart and screenshotHow my own router splits the work