The AI, and what it costs
Last updated 27 October 2026 · the measured figures are read live
The short version. Every GreenSpecs scan is read by Google's Gemini. That costs energy, water and carbon, and the cost is ours, not something we can hand to someone else. Below: what the model does here, what a scan costs us in money, what we estimate it costs the planet, and what we do to keep that number down.
We do not claim a scan pays for itself. We cannot measure a purchase that did not happen.
What the AI does here
| Call | Model | When | What it returns |
|---|---|---|---|
| Identify | Gemini 2.5 Flash Lite | Photo scans only | The product and brand, in about a second, so the result screen is not a spinner. It never touches the score. |
| Score | Gemini 2.5 Flash | Every scan that is not served from cache | The five signals, the Material Score, the written reading, and the confidence. It thinks before it answers. |
| Facts panel | Gemini 2.5 Flash | Every scan that is not served from cache | The panel's values and their source tiers. Runs alongside the score call, not after it. |
| Swaps, compare, Ask Sustainability, battles | Gemini 2.5 Flash | Only when you tap them | Alternatives, a head-to-head, an answer, a round. None of these run on their own. |
So a typed scan is two model calls and a photo scan is three. Nothing in the app calls a model on a timer, in the background, or to fill a screen you did not ask for.
Several things people assume are AI are not. The letter-grade bands are in code. The confidence step-down past 180 days is in code. Category averages, deltas and percentiles are precomputed on a weekly schedule, never per scan and never by the model. Certification claims are checked against the issuer's own register, and B Corp by a person. And the model is not allowed to mark a number verified: that gets downgraded to estimated in code before it is stored.
What a scan costs us
| Measured | Figure |
|---|---|
| Scans so far | 1,358 |
| Of those, answered from cache with no model call | 140 |
| Scans that called the model | 1,218 |
| What we have paid Google for all of it | $6.67 |
| Average per scan that called the model | $0.0055 |
Read from our production database when this page was built, at most five minutes before you loaded it. We do not keep a second, nicer set of numbers.
What a scan costs the planet — our estimate
We cannot meter Google's data centre from here, so this is an estimate built from Google's own published figures, and we would rather show the arithmetic than a single confident number.
Google's August 2025 technical paper puts the median Gemini text prompt at 0.24 Wh of energy, 0.26 mL of water and 0.03 g CO2e. That is its wider "comprehensive" boundary, which counts idle capacity and data-centre overhead as well as the chips doing the work. The narrower active-chips-only boundary is 0.10 Wh. We use the wider one deliberately: it is the number that counts the whole machine.
A GreenSpecs scan is not a median text prompt. It is two or three calls; one of them carries a photo and a rubric a few thousand tokens long; and the score call thinks before it answers, which burns as output. Google does not publish a per-call figure for an image prompt with thinking. So we count a scan as 4 to 12 median prompts — the low end a little above three calls at median, the high end allowing for the photo, the long prompt and the thinking tokens all landing heavy.
| One scan that called the model | Estimate |
|---|---|
| Energy | 1.0 to 2.9 Wh |
| Water | 1.0 to 3.1 mL |
| Carbon | 0.12 to 0.36 g CO2e |
| Everything GreenSpecs has scanned, so far | Estimate |
|---|---|
| Energy | 1.17 to 3.51 kWh |
| Water | 1.27 to 3.80 litres |
| Carbon | 0.146 to 0.438 kg CO2e |
For scale, the upper end of one scan is about three minutes of a laptop charger, and less than a teaspoon of water. That is not nothing, and it is not an emergency. It is a real cost with our name on it.
What we do to keep it down
- A typed scan of something already scanned calls no model at all. The result is reused from the database, and the row is marked as served from cache. 140 of 1,358 scans so far needed no model call because of this. Photo scans skip the cache, because the photo may be of something else.
- The fast "what am I looking at" call uses the smallest model, Flash Lite, with no rubric and no thinking. The heavy model is only asked the hard question.
- Comparisons are precomputed weekly, by one scheduled job, instead of being worked out again on every scan.
- The two heavy calls run in parallel, so the app does not retry out of impatience and double the cost.
- Your last 24 scans are cached on your device, so re-opening something you just scanned calls nothing and sends nothing.
- There is a hard daily ceiling on model spend. Past it, scans stop calling Gemini and the app says so until the next day. It exists to stop a runaway loop, and it caps the energy as well as the bill.
- Photos are not kept. A photo is sent to be read and then dropped, so we are not also storing it forever.
- Nothing is generated speculatively. No prefetched answers, no background rescoring of things nobody asked about, no model call to draw a screen.
What we will not claim
- That a scan is carbon negative, climate positive, or pays for itself. One avoided purchase almost certainly outweighs the scan that prompted it — a single bottle of detergent carries far more embodied carbon than a few watt-hours — but we cannot observe a purchase that did not happen, so we will not put a number on it.
- That the estimate above is precise. It is a range from someone else's median, and the range is wide for a reason.
- That Google's figures are independently audited. They are self-reported, not peer-reviewed, and critics have noted the paper does not publish total prompt volumes. We use them because they are the most specific figures the provider has published, and we will replace them when something better exists.
- That our offsets make this zero. We do not buy any, and we would rather cut the calls than buy a certificate.
When the AI is wrong
It is sometimes wrong. It reads a label in a photograph, in whatever light the aisle has, and the parts of a product that matter most are usually not printed on the pack at all. Every result carries a confidence for that reason, and the written reading names what the label does not say.
If a result is wrong, rate it in the app or email hello@greenspecs.app. A person reads every report, and a product can be queued for a re-score, which replaces the stored result. On the Facts panel, a verified figure replacing an estimate is logged separately, with both values, the source, and the score before and after.
Where the numbers come from
- Google Cloud, Measuring the environmental impact of AI inference, August 2025: the published figures and method.
- Our own spend, scan counts and cache hit counts: our production database, read when this page was built.
- The model names and the number of calls per scan: the GreenSpecs worker source, described above exactly as it runs.
What is next
Two things we intend to publish rather than estimate: the real token counts per scan, which we already record, and a per-scan figure for an image prompt with thinking if Google publishes one. If either lands, this page changes and the date at the top moves.