TokenAtlasProduct copy generation
10K SKUs × 600 output tokens × $10/1M (GPT-4o) = $60 one-time. With Sonnet: $90. With Haiku: $7.50. Run once, edit by hand.
Semantic search
Indexing 50K products with text-embedding-3-small: $1 one-time. Query cost per search: ~$0.00002. Even at 1M searches/mo: $20. Embeddings are nearly free.
Support automation
50K tickets/mo deflected by AI: ~$300 on GPT-4o-mini, ~$60 on Haiku, ~$15 on Flash. ROI obvious vs human agent cost.
Choosing a model by SKU tier
Blended at 70/30: GPT-4o = $4.75/1M, Claude 3.5 Haiku = $1.76/1M, DeepSeek V3 = $0.519/1M. On 100,000 long-tail SKUs at 200 output tokens each (20M output tokens, assume negligible input for a templated prompt), Haiku models to roughly $80 and DeepSeek V3 to roughly $22 — worth testing against GPT-4o's output-only cost of $200 for SKUs where a top-tier model isn't needed.
What moves the estimate most for catalog workloads
Catalog size is the dominant driver since it multiplies every per-SKU cost directly, and it is also the input most likely to grow. Output length per description ranks second — a template that asks for 150 words instead of 300 roughly halves the output-token cost. Re-running the same catalog size across models is the fastest way to see where the trade-off actually sits.
Reading the estimate: what it is and isn't
A catalog cost estimate models the token volume of generating or searching descriptions once at the stated size and rate — it does not account for re-generation when a SKU is edited, or for support tickets referencing a product. Add a re-generation rate assumption (e.g. 10% of SKUs updated monthly) if that is part of the real workflow, since it changes the effective monthly volume.
A second scenario: seasonal catalog refresh
A retailer refreshing 30,000 seasonal SKUs at 600 output tokens each on Haiku (output rate $4/1M) models to 18M output tokens x $4/1M = $72, versus $180 on GPT-4o's $10/1M output rate for the same volume — run this alongside the steady-state 10K-SKU figure on this page since seasonal spikes are a separate line item, not a scaled version of the baseline.
Translating catalog cost into cost per order
Dividing total modelled AI spend (copy + search + support) by monthly orders gives a per-order figure comparable against fulfillment or payment-processing fees: at $60 + $20 + $300 = $380/month against 5,000 orders, that's $0.076/order — a number worth tracking alongside other per-order costs rather than looking at AI spend in isolation.
Assumptions checklist for a catalog estimate
— Active SKU count and expected monthly re-generation rate, not just total catalog size — Average output length per description and per support reply — Monthly search query volume and average ticket volume — Whether embeddings are re-indexed on every catalog update or only on net-new SKUs
Frequently asked questions
- Which model for product descriptions?
- GPT-4o or Sonnet for hero SKUs; Haiku/Flash for the long tail.
- Is RAG worth it for product search?
- Yes — hybrid (BM25 + embeddings) beats keyword by 20–40% on conversion in published e-commerce studies.
- What about image generation cost?
- Lifestyle imagery via image models: $0.02–$0.08 per image. Plan accordingly.