Every number we have published
41 figures from our own production work, each with what it was measured from and the article that reports it. Nothing here is a vendor price list, and nothing here is an estimate we have not labelled as one.
One rule governs this page. A figure appears here only once it has appeared in a published article. The index never discloses a number an article does not, and never runs ahead of something we have not written yet. Where we hold a number and no article has reported it, it is absent on purpose.
Values are reproduced in the units the articles use. Some are in CNY because CNY is the currency we were billed in, converted at ¥6.72 to the dollar where a dollar figure is shown. We do not re-derive them here, because that would create a second version of the same number.
How each figure was obtained
| Basis | Meaning |
|---|---|
measured | from records or exports (invoices, account reports) |
recalled | reconstructed from memory of the work, not reconciled |
mixed | some figures measured, some recalled |
derived | arithmetic on the rows above it |
Cost 23
| What was measured | Figure | Basis | When | Source |
|---|---|---|---|---|
| Generation cost, one five-shot 38.7-second vertical adInvoice line for a single submitted job, at ¥6.72 to the dollar. | $0.51 (¥3.42) | measured | 2026-08 | What a 38.7-second AI ad actually costs |
| Generation cost per second of finished videoGeneration total divided by 38.7-second runtime. The list rate was ¥0.09 per second. | $0.0131 (¥0.088) | measured | 2026-08 | What a 38.7-second AI ad actually costs |
| Billed rate at 768P, current modelThe rate actually invoiced, not a published price list. | ¥0.09 / second | measured | 2026-08 | What a 38.7-second AI ad actually costs |
| Cost to regenerate one rejected shotOne shot regenerated with an amended prompt, billed as its own line. | $0.067 (¥0.45) | measured | 2026-08 | What a 38.7-second AI ad actually costs |
| Five-shot ad including one rejected shotFour shots at ¥1.80 ($0.27) plus one regeneration at ¥0.45 ($0.067). | $0.33 (¥2.25) | measured | 2026-08 | What a 38.7-second AI ad actually costs |
| Human editing labour, per finished itemOur own team’s pure labour cost per finished edited item, excluding media spend. | ¥15–20 | measured | 2026 | What a 38.7-second AI ad actually costs |
| Generation cost, one 200-shot batch (earlier route)Production accounting for the batch as recalled by the team; no reconciled ledger exists. | ¥400 (≈$60) | recalled | 2026-07 | Cost per AI ad is a denominator problem |
| Generation cost per approved finished item (earlier route)Batch cost divided by the items that passed review. A generation line, not a production cost. | ≈¥1.05 (≈$0.16) | recalled | 2026-07 | Cost per AI ad is a denominator problem |
| Share of built items passing review (earlier route)Pass rate across the batch. The review labour this implies is not costed in the figure above. | ~95% | recalled | 2026-07 | Cost per AI ad is a denominator problem |
| Uses per generated shot (earlier route)800 shot-uses divided by 200 generated shots. Exact by construction: two slots per item. | 4.0 | recalled | 2026-07 | Cost per AI ad is a denominator problem |
| Pairings available from one batch100 A-type slots multiplied by 100 B-type slots. | 10,000 | derived | 2026-07 | Cost per AI ad is a denominator problem |
| Share of the pairing space actually builtAbout 400 finished items against 10,000 available pairings. | ~4% | recalled | 2026-07 | Cost per AI ad is a denominator problem |
| Rate, the earlier model used on that routeBilled rate on the earlier route. More than twice the current rate per second. | ¥0.2 per generated second | recalled | 2026-07 | Cost per AI ad is a denominator problem |
| Spread across three purchase channels, same modelStructured evaluation of three channels. Identical output; only the supplier changed. | ¥0.11 → ¥0.83 / second (7.5×) | mixed | 2026-08 | The same AI video model at 7.5× the price |
| Cheapest channel, 768P and 2KThird-party reseller. Carries a manual consistency step that the invoice does not show. | ¥0.11 / s (≈$0.016) · ¥0.18 / s (≈$0.027) | mixed | 2026-08 | The same AI video model at 7.5× the price |
| Consumer app channel, 768P and 2KSecond of the three channels compared on identical output. | ¥0.43 / s (≈$0.064) · ¥0.70 / s (≈$0.104) | mixed | 2026-08 | The same AI video model at 7.5× the price |
| Vendor hub channel, 2KMost expensive of the three channels. Six-minute generation, and still misreads percentage signs. | ¥0.83 / s (≈$0.124) | mixed | 2026-08 | The same AI video model at 7.5× the price |
| Channel gap on one 38-second videoCheapest against most expensive channel at 2K over 38 seconds. | ¥27.36 (≈$4.07) | derived | 2026-08 | The same AI video model at 7.5× the price |
| Channel gap over 1,000 generationsThe same 38-second gap scaled to a thousand videos. | ¥27,360 | derived | 2026-08 | The same AI video model at 7.5× the price |
| Saving from the cheap channel on a 12-second shot¥0.72 per second saved over 12 seconds. Ten minutes of editor time costs more than this. | ¥8.64 | derived | 2026-08 | The same AI video model at 7.5× the price |
| Price range across ten AI video platformsStructured evaluation of ten platforms, using published or quoted rates rather than negotiated prices. | ¥0.11 → ¥1.90 / second (17×) | mixed | 2026-08 | The same AI video model at 7.5× the price |
| Same channel recorded at two different ratesBilling records against the later evaluation. Either the rate or the credit package changed; we do not know which. | ¥0.09 vs ¥0.11 (768P) · ¥0.15 vs ¥0.18 (2K) | measured | 2026-06 to 2026-08 | The same AI video model at 7.5× the price |
| Image generation cost, five vocabulary sheetsMeasured baseline for this endpoint and quality setting. The billed figure is not published. | ≈$0.03 | measured | 2026-09 | Why AI images get the text wrong |
Performance 10
| What was measured | Figure | Basis | When | Source |
|---|---|---|---|---|
| Creatives compared in one live accountOne direct-response advertiser account, same buying system and conversion event for both pools. | 26,013 AI vs 25,080 human | measured | 12 weeks | AI vs human ad creative: 26,013 creatives |
| Share of spend against share of conversions, AI poolHow much AI creative over- or under-delivered relative to the budget it absorbed. | 12.98% spend vs 12.96% conversions (+0.03pp) | measured | 12 weeks | AI vs human ad creative: 26,013 creatives |
| Cost per conversion over the full periodIndexed to human creative. The two pools came out level across the twelve weeks. | Level (human = 100) | measured | 12 weeks | AI vs human ad creative: 26,013 creatives |
| Budget received per creativeAI creative received about a seventh of the budget per creative that human creative received. | 14.4 (human = 100) | measured | 12 weeks | AI vs human ad creative: 26,013 creatives |
| Cost per conversion by four-week block (human = 100)The reversal the twelve-week average hides: roughly 4% cheaper, then roughly 6% more expensive. | 96 → 103 → 106 | measured | 12 weeks | AI vs human ad creative: 26,013 creatives |
| Weekly AI output across the periodOutput rose about thirtyfold while the workflow was also changing. | ~150 → ~5,000 creatives | measured | 12 weeks | AI vs human ad creative: 26,013 creatives |
| Day-1 retention rate, AI against humanA single week whose report carried downstream behavioural metrics rather than conversions alone. | ~16% lower (relative) | measured | one week of the period | AI vs human ad creative: 26,013 creatives |
| Cost per key action, AI against humanSame single week. Key-action rate was about 4% lower and cost per key action correspondingly higher. | ~17% higher | measured | one week of the period | AI vs human ad creative: 26,013 creatives |
| Cost per conversion in that same weekNot the twelve-week figure. In this week AI bought conversions at the same price but of lower quality. | ~12% worse | measured | one week of the period | AI vs human ad creative: 26,013 creatives |
| Creatives recording any spend at allA larger share of AI creative received no delivery whatsoever. | 72% AI vs 83% human | measured | month-level cut | AI vs human ad creative: 26,013 creatives |
Reliability 8
| What was measured | Figure | Basis | When | Source |
|---|---|---|---|---|
| Image cells checked by handFull enumeration at full resolution with 2–8× enlargement of suspect cells. No sampling. | 540 (270 per pipeline) | measured | 2026-09 | Why AI images get the text wrong |
| Hard content errors, model-written against code-written labelsA wrong Chinese character and a mislabelled animal on the model-written side. | 0 / 270 vs 2 / 270 (0.74%) | measured | 2026-09 | Why AI images get the text wrong |
| English spelling correctThe failure was not English. It was concentrated in Chinese labels. | 270 / 270 both sides | measured | 2026-09 | Why AI images get the text wrong |
| Chinese labels correctCode-written labels cannot misspell; model-written labels can. | 270 / 270 vs 269 / 270 | measured | 2026-09 | Why AI images get the text wrong |
| What three deterministic gates changed against human reviewRegressed against thirty historical sheets. The gates cost nothing to run. | Caught 2 missed defects; overturned 3 false alarms | measured | 2026-09 | Why AI images get the text wrong |
| Near-white ratio, broken dividers against healthy sheetsThe pixel measurement that separated two genuinely thin dividers from healthy ones. | 0.68 and 0.49 vs 0.95–1.00 | measured | 2026-09 | Why AI images get the text wrong |
| Rebuilt pipeline, first attemptFive sheets end to end with zero text errors of any kind. | 5 / 5 clean, 0 retries, mean 19.0s | measured | 2026-09 | Why AI images get the text wrong |
| Delivery size for the same 270 cellsReal and reproducible, but not attributable to the variable under test. | 48 MB (JPG) vs 154 MB (PNG) | measured | 2026-09 | Why AI images get the text wrong |
Where these come from
- What a 38.7-second AI ad actually costs
- Cost per AI ad is a denominator problem
- The same AI video model at 7.5× the price
- Why AI images get the text wrong
- AI vs human ad creative: 26,013 creatives
What this page does not claim
A figure is only as good as the sample behind it, and several of these are small. The twelve-week account comparison is one client. The 38.7-second bill is one job. The 540-cell image comparison moved six variables and attributes one. Each article states its own limits in a Method section, and those limits are part of the number.
Prices also expire. Model rates, channel quotes and platform pricing change within months, so each cost row carries the date the work was done. If you are planning against a figure here, ask us for a current one first. We re-run the test before relying on it ourselves.
If a number here is wrong, tell us. Corrections are published with a date and the original figure kept visible.