One real invoice: $2.2k a month of LLM, and what compression gets back
An Azure bill from a B2B SaaS, read line by line. A 10 % cache-hit rate, one LLM call per candidate, a timestamp at the top of the system prompt — and an honest estimate of what a compressor alone recovers.
A B2B SaaS shared its February 2026 Azure invoice and read access to its codebase. Foundry Models: $2,218. Of that, uncached input $1,140 for about 912M tokens, output $683 for 68M, and cached input $13 for 106M tokens. That last line is the one to read first: a cache-hit rate around 10 %, down from 21 % in January. You can read it straight off the invoice, no application log needed.
Where the tokens come from
- Candidate matching: one LLM call per candidate, 52 per search, in parallel, on a model without caching.
- Document adaptation: one call per work experience, 43 per session.
- Document generation: 13 parallel calls, each with a variable system prompt placed before the identical résumé — so no shared prefix.
- Document analysis: one call per page on the largest model.
- Summaries with reasoning effort set to high.
- Chat: the date to the minute, memories and tool context at the top of the system prompt, ahead of 10k tokens of stable instructions — the cache only serves steps of the same turn.
The application's own usage table captured about 20 % of the volume and priced one model 13× too high, so the admin dashboard attributed 70 % of spend to matching. It is closer to 17 %.
What each lever is worth
| Lever | Estimated saving / month | Quality risk |
|---|---|---|
| Cacheable prefix (reorder system prompt, freeze timestamps) | $400–600 | none |
| Batching fan-out calls | $100–200 | none |
| Output and reasoning settings | $130–270 | low |
| Structural JSON compaction (lossless) | $100–135 | none |
| Neural text compression at −20 % | $20–50 | low |
On this invoice a compressor alone is worth $20–50 a month. The cache and the fan-out are worth $600–1,000. That is why the Tokeen SDK ships prefix ordering, batching and JSON compaction next to the compressor — and why the first thing we do is read your invoice.
The five-minute qualification
Two lines on any provider invoice — cached input and uncached input — give the hit rate. Below 30 % on a multi-turn product, the prefix is unstable and the fix is structural. Above 80 %, compression on cold prefixes is the remaining lever. Send us yours; we answer with a number.