AI Tagging Cost Guide. Library Sizes 1k → 1M

A note on the numbers. Every price below comes from Google's public Gemini API pricing page, checked on 24 September 2026. Every token count is either from Google's own token documentation or an illustrative assumption that is labelled as one. Model prices change often, so treat the worked examples as a method with today's inputs, not a quote, and re-run them with the AI tagging cost calculator using whatever rate your provider publishes on the day you read this.

Two-stage tagging explained

A vision model can only tag products it has been told about. The prompt therefore carries two things: the image, and a list of candidate products from your catalogue with enough detail to tell them apart. The image cost is fixed by resolution. The list cost grows with every product you include, so a 4,000-SKU catalogue pasted into every call would dwarf the image itself and also make the model worse at choosing. Two-stage tagging exists to fix that.

How Idukki runs it

Idukki's tagger follows this shape, and it is worth being exact because some older descriptions of it are out of date. It runs on Google's Gemini vision models (not Claude). For a business with up to 150 active products, every product goes into the prompt. Above that, Idukki embeds the post's caption and labels and sends the 60 closest products instead of the whole catalogue, falling back to a capped list if the post has no usable text. Where product attributes such as colour family and style have been derived, they are added to each product line so the model can separate near-identical titles. By default, matches below 0.5 confidence are dropped, matches from 0.5 up to 0.75 land in a review inbox alongside caption-based suggestions, and matches at 0.75 or above are tagged automatically. Each image can carry up to 20 matches. For video posts the tagger reads the thumbnail frame rather than every frame, which keeps video no more expensive to tag than a photo. The mechanics of the full pipeline are covered in the tagging stage of the AI UGC loop, and the product surface is on the AI tagging page.

Cost at scale: the worksheet

Per-asset cost is input tokens times the input rate, plus output tokens times the output rate. Output includes any "thinking" tokens the model spends before answering, which Google bills at the output rate, so a model that reasons at length before writing a ten-token JSON answer can cost more on the output side than the input side. Fill in the five inputs below for your own library.

With those values, one asset uses about 2,232 input tokens and 400 output tokens. On Gemini 3.6 Flash at standard rates that is roughly $0.00167 of input plus $0.0015 of output, so about $0.0032 per asset, or $3.17 per 1,000. Google's batch tier charges half the standard rate for the same model, which suits a one-off backfill where nobody is waiting on the result. Google's page also states that Gemini 3.6 Flash rises to $1.50 in and $7.50 out per million tokens from 1 January 2027, so any budget that runs into next year should use the higher figure. Gemini 3.5 Flash-Lite is cheaper again at $0.30 in and $2.50 out (standard), worth testing if its accuracy holds on your catalogue.

What actually moves the bill

• Catalogue lines in the prompt. Sending 150 products instead of 60 roughly doubles prompt tokens in this example. Shortlisting is the single biggest saving on a large catalogue.
• Image resolution. A 1080 px square is four tiles; a 384 px thumbnail is one. Products that fill the frame often tag fine from a smaller image, so test a downscaled sample before paying for full resolution.
• Thinking tokens. If your model supports a thinking budget, cap it. A matching task with a fixed list rarely needs long reasoning, and every thinking token is billed at the output rate.
• Frames per video. Tagging five sampled frames costs five images. The cover frame is usually enough for a product-led clip; multi-product try-ons may justify more.
• Re-runs. A catalogue change does not require re-tagging the whole library. Re-tag only posts whose shortlist would change, or new posts, and the steady-state cost is just your monthly intake.

The embedding step for the shortlist is small by comparison. A caption and label set is typically a few hundred text tokens, and Google lists its current embedding model at $0.20 per million text tokens, so embedding 1,000 posts of about 300 tokens each costs around six cents. It matters for architecture, not for budget.

When to keep humans in the loop

No vision model gets every product right on every photo, and nobody can give you a trustworthy universal error rate for yours: accuracy depends on how distinctive your products are, how cluttered customer photos are, and how similar your SKUs look to each other. A skincare range in near-identical bottles is harder than a furniture catalogue. The honest approach is to measure it on your own library and design the review queue around the result.

Size the review queue before you size the model bill

The review queue is where the money goes. As an illustrative calculation: if 20% of a 100,000-asset backfill lands in review and a reviewer spends 10 seconds per suggestion, that is 20,000 decisions and about 56 hours of work. At almost any loaded hourly rate, that is several times the model bill in the table above. Halve the review share and you save more than switching to the cheapest model would. So the order of work is: shortlist well, set thresholds from evidence, and only then shop around on per-token price.

1. 1Label a sample. Take 200 to 300 recent posts, have one person tag them by hand, and treat that as ground truth.
2. 2Run the tagger on the same sample. Record every suggested product with its confidence score.
3. 3Plot precision by band. For each 0.05 confidence band, what share of suggestions were right? The band where precision reaches the level you would accept on a live PDP is your auto-tag threshold.
4. 4Count the middle. The share of assets whose best match sits between your minimum and auto-tag threshold is your review rate. Multiply by library size and seconds per decision.
5. 5Repeat quarterly. New product lines and seasonal content shift the numbers, so the thresholds that were right in spring may not be right by peak season.

For the review itself, keep it fast: show the post, the suggested product image and title side by side, and make accept and reject single actions. The broader moderation workflow (brand safety, rights, sentiment) is a separate queue with different reviewers, covered in moderating UGC at scale. In Idukki, suggestions from both vision and captions surface together for accept or reject, and posts can still be tagged by hand, with hotspots, as described in the manual tagging help article.

Build it yourself or use a platform

The token bill is the smallest part of an in-house tagger. The real costs are the catalogue sync that keeps product ids current, the embedding index and its re-builds, the queueing that stops a backlog spiking your API spend in one night, the review interface, and the on-call time when a provider changes a model or a rate. None of those shows up in a per-token estimate. Idukki prices on widget impressions and includes AI tagging in every plan according to its pricing page, so there is no separate per-asset tagging bill to forecast. For the fuller build-versus-buy arithmetic, see the cost of UGC: build versus buy.