Cost scenario · Content
Generating content at scale
Product descriptions for a catalogue, or listing summaries, or short marketing copy. A short brief goes in and several hundred words come out.
This is the inverse of document processing: output dominates, which makes it the most expensive shape per token of anything on this site.
The shape of one piece
A 600 token prompt carrying tone rules and two examples, constant, plus around 600 tokens of product facts. Output is roughly 700 words. The 1.3 calls per piece is the regeneration rate — some output gets rejected and rerun.
What it costs per month
Computed from the verified rates in the price table, last reviewed August 5, 2026. Figures assume caching is enabled — the uncached column shows what skipping that flag costs.
| Volume | Ministral 3 8BCheapest that fits | Magistral MediumMid tier | Claude Fable 5Frontier |
|---|---|---|---|
| Small catalogue5,000 pieces / month | $1.58$1.58 uncached | $34.50$34.50 uncached | $258.00$285.00 uncached |
| Mid catalogue100,000 pieces / month | $31.50$31.50 uncached | $690.00$690.00 uncached | $5,160.00$5,700.00 uncached |
| Marketplace1,000,000 pieces / month | $315.00$315.00 uncached | $6,900.00$6,900.00 uncached | $51,600.00$57,000.00 uncached |
On this shape the spread between the cheapest model that fits and the frontier one is 181.0×, which is the largest single decision on this page. Enabling prompt caching removes a further 9% on Claude Fable 5, because 50% of the input on this workload is a stable prefix.
Where a provider publishes no cached input rate, the cached and uncached figures above are identical. That is a gap in what the provider discloses, not a model that cannot cache.
What actually drives the bill
Output is the expensive side
Output typically costs three to five times input on the same model, and here it is nearly half the tokens. Every extra sentence of length is billed at the premium rate on every piece.
The regeneration rate
Nobody accepts every generated piece. A thirty percent rejection rate is a thirty percent cost increase, and it never appears in a budget built from the published price per token.
Length creep
Without an explicit cap, generated copy grows. A prompt that produced 500 words in testing produces 800 in production on richer inputs, and that is a sixty percent output increase nobody decided.
How to cut it, in order of return
- Set a hard word count and enforce max_tokens20–35% of output
Two mechanisms, both needed. The word count aims the model, max_tokens bounds the damage when a rich input makes it ramble.
- Fix the prompt instead of regeneratingthe whole rejection premium
A thirty percent rejection rate is a prompt problem being paid for with tokens. Ten rejected pieces read carefully will usually identify one missing constraint.
- Batch everything except live previewsabout 50%
Catalogue generation is a background job by nature. Only the piece someone is watching generate needs a synchronous call.
- Cache the tone and example block20–25%
Half the input is constant across every piece in the catalogue.
What people get wrong budgeting this
- The rejection rate is the number that decides this budget and it is never in the vendor estimate. Measure yours before committing to volume.
- Output converges at scale. A thousand descriptions from one prompt share a structure, and fixing that usually means a longer prompt — which costs more per piece.
- Regulated categories need human review regardless of quality, and that review is the real cost of the workload.
Price your own version
The shape above is representative, not measured from your system. Put your real prompt and volume through these and the number becomes yours.
- LLM API Cost CalculatorRequests per month in, dollars out. Input, cached input and output priced separately.
- max_tokens PlannerA ceiling derived from your p95, checked against window and quota.
- Batch API Savings CalculatorThe 50% discount most teams never claim, priced at your volume.
- Prompt Cache Structure CheckerFinds cache-busting content and measures the prefix that survives it.
Before committing to volume, the cost control checklist is ordered by saving per hour of effort, and every rate used here links to the provider page it came from on the Mistral AI and other provider pages.