Examples

Five real runs.
Nothing cleaned up.

Five jobs, handed to Else one after another on 3 October 2026. Below is what we asked for, what came back, and what each run cost. All five produced a file you can open, and every one of them has a replay link you can watch without an account. One of the five wrote its file but hit an error registering the delivery; that is written up as it happened, not tidied away.

5
runs, back to back, on one afternoon
5
finished files, linked below and unedited
1
delivery that errored while being registered
$2.526918
total across all five runs, added up
Slide deck Finished · result verified

“Build me a 6-slide HTML deck I can open in a browser, titled "Why our support queue is slow". Use only these numbers: 4,120 tickets last quarter; median first reply 9h 40m; 38% of tickets are password resets; 2 agents on weekends vs 6 on weekdays; satisfaction 3.4 out of 5. Include one chart slide and one recommendations slide. Save it to /outputs as a single self-contained .html file.”

What it did
  1. Showed the plan before starting: a responsive six-slide deck with keyboard navigation, a slide counter and a dark theme, then a named slide for each of the five figures we supplied.
  2. Wrote the file in the secure workspace, rendered a preview of it, and read it back to verify before handing it over.
  3. Finished in 1m 00s and reported what it had built, slide by slide.
What we got

Six slides: a title and summary card, a queue-performance overview, the 38% password-reset bottleneck, the weekend-versus-weekday staffing gap, a chart slide carrying a donut of 38% password resets against 62% other tickets and a bar chart of six weekday agents against two on weekends, and a recommendations slide that lands on self-service password reset and weekend shift rebalancing. Keyboard navigation, indicator dots and Next/Previous controls were not asked for; it added them anyway. It also wrote out a PNG preview of the deck alongside the file.

why_our_support_queue_is_slow.html 19.8 kB · opens in a new page

Watch the replay

$0.205452 Fast tier 1m 00s · finished on its own
Report Finished · result verified

“We got three quotes to re-roof a 92 sq m house and I can't hold them in my head. Quote A: 14,200 pounds, slate, 12-year workmanship guarantee, can start in 3 weeks, scaffolding included, 40% deposit. Quote B: 11,650 pounds, concrete tile, 5-year guarantee, can start next week, scaffolding charged separately at about 1,400 pounds, 25% deposit. Quote C: 16,900 pounds, slate, 20-year insurance-backed guarantee, earliest start is 9 weeks away, scaffolding and waste removal included, no deposit until work starts. Build me a decision table comparing true total cost, risk and timing, then say plainly which one you'd pick and what you'd ask each of them before signing. Save it to /outputs as a markdown file.”

What it did
  1. Said up front it would compare on landed cost rather than headline price, then on risk and timing, then give a straight pick and a question list.
  2. Worked the arithmetic out in the open before writing anything — each quote's true total, its cost per square metre over the 92 sq m roof, and the cash you are exposed to before work starts.
  3. Finished in 1m 40s with the file written and a recommendation stated in the conversation.
What we got

A seven-section decision document, a pick, and nine questions to put to the three builders — three each, written for that quote. It found the thing we hadn't: adding scaffolding back turns Quote B's £2,550 headline advantage into £1,150, so the slate roof costs £1,150 more than the concrete-tile one rather than £2,550 less. It picked Quote A on the condition that the 40% deposit comes down, called £5,680 before anyone sets foot on site an uncomfortable risk, and named the standard it was measuring against. Where a cost was not stated it flagged it to ask rather than assuming, warned that concrete tiles are heavier and the roof timbers may need checking, and noted that an insurance-backed guarantee survives the company ceasing to trade while a workmanship one does not. It also kept two escape hatches in plain sight: pick B if water is coming in now, pick C if you want zero deposit risk and will not negotiate.

roofing_quotes_comparison.md 7.8 kB · plain text

Watch the replay

$0.458311 Fast tier 1m 40s · finished on its own
Email Finished, start to end

“Write the email I've been putting off. I have to tell a freelance designer we're not renewing her contract after 14 months. The real reason is our budget was cut, not her work — and I genuinely want to hire her again next year. Under 200 words, warm but unambiguous, no corporate filler, no "synergy", no apologising three times. Save the final version to /outputs as a markdown file.”

What it did
  1. Stated the three things it was going to do — draft under 200 words, save it as markdown, hand the file back — and then did them.
  2. Wrote the email in one pass, with a subject line, and saved it.
  3. Finished in 24 seconds, the shortest and cheapest run on this page.
What we got

A 184-word email with a subject line, sendable after three fill-ins. The decision is in the second sentence and attributed to the budget, not to her: “this is purely a financial decision and is in no way a reflection of your work.” It apologises nowhere. The rehire line is bounded to something you can actually keep — it promises to make contact when the next budget cycle opens, not to re-sign her. It closes by asking what hours and deliverables are still outstanding, which is the practical question the email we had been putting off would have forgotten.

contract_non_renewal_email.md 1.0 kB · plain text

Watch the replay

$0.079912 Fast tier 24s · finished on its own
Research Finished · result verified

“Compare the three best-reviewed countertop air fryers under $150 sold in the US right now. Use current sources and cite them. Give me a comparison table with capacity, wattage, price, and the single biggest complaint reviewers make about each one. Save the finished comparison to /outputs as a markdown file.”

What it did
  1. Turned on the Web research add-on and went looking for current test-and-review coverage rather than retailer listings.
  2. Read across five publications — Wirecutter, Business Insider, Bon Appétit, CNET and RTINGS — and settled on the three models they agree on.
  3. Wrote the comparison out with its citations and finished in 2m 31s.
What we got

A comparison of the COSORI TurboBlaze 6-quart, the Instant Vortex Plus 6-quart ClearCook and the Ninja Air Fryer Max XL, each with capacity, wattage, MSRP against street price, citations, and the one complaint reviewers keep making: confusing presets that lock the fan speed on the COSORI, a chemical smell through the first several uses on the Instant, and a round basket that wastes flat cooking area on the Ninja. This is the most expensive run on the page at $1.116063, and it is the shape of job that costs the most — reading real pages across the open web is slow and it is charged by what it reads, not by what it writes.

air_fryer_comparison.md 7.1 kB · plain text

Watch the replay

$1.116063 Fast tier 2m 31s · finished on its own
Research Answered · delivery registration errored

“Short, tight research job — keep it cheap. Compare renting vs buying a home broadband router in the UK. Search the web, open at most two pages, and cite what you use. Give me a markdown table with typical monthly rental cost, typical one-off purchase cost, and the main catch of each option, plus a two-line verdict. Save it to /outputs as a markdown file.”

What it did, and what went wrong
  1. Read the instruction back correctly, including the two-page cap, and laid out four steps: search, open up to two sources, draft the table and verdict, save the file.
  2. Used two sources and cited both by name — Choose.co.uk and Which? — and wrote the comparison and verdict.
  3. Wrote the file, then hit an error on the way out. Registering the delivery failed, so the run ended at 1m 37s with the report in the conversation and no file attached to it.
What we got

The work was delivered; the delivery was not registered. The file was written to the workspace and the full report — the rental-versus-purchase table, the catch on each option, the two-line verdict and both citations — was printed in the conversation, so nothing was lost. But Else said so itself rather than quietly dropping it: “The document was written to /outputs/router_rent_vs_buy_uk.md, but workspace file verification encountered a system error preventing deliverable registration in this turn. The complete output is provided above for your direct access.” That is a fault on our side, not a model refusing to answer, and we were still charged $0.667180 for the run. The file below is that file, pulled out of the workspace unedited; the replay shows the error notice exactly as it appeared.

router_rent_vs_buy_uk.md 2.1 kB · plain text

Watch the replay

$0.667180 Fast tier 1m 37s · answered, delivery not registered
How to read this page

Every number here came off a receipt.

Not a model, not an average, not a forecast. Each amount is the figure Else itself printed at the end of that run, next to the engine tier that produced it — all five of these came back on Fast. A different tier gives a different receipt, which is why an amount on this site never appears without the tier beside it.

The five files are byte-for-byte what Else wrote. We did not edit, reformat, restyle or trim them, and we did not tidy up the run whose delivery errored on the way out. If something went wrong, it says so where it went wrong.

These five are not the only runs on the site. Each one now has a replay link, so you can open the run itself and read it without an account — the replay shows the conversation, the tier and the files, though not the amount. They are still not in the downloadable cost dataset: that dataset is built from the gallery entries, and these five are not gallery entries. The twelve runs itemized at what AI tasks cost all ran on Smart, where these five ran on Fast, so the two sets are not directly comparable on price.

Hand Else your first job.

Try Else
Try Else