The same tape that answers your auditor turns out to answer your CFO. Because every recorded model call carries its token counts, Wytness can show you every model your agents used this cycle, the tokens they spent, and an estimate of what that cost you at your LLM vendor — per model, from evidence rather than vibes.
First, what this is not
This is not a Wytness bill. We don't charge for model usage and we never see your vendor invoice. Every dollar on the Models page is an estimate of your spend with Anthropic, OpenAI, or whoever you call — computed from the tokens your SDK captured, priced at published rates. Your real invoice will differ wherever committed spend or volume discounts apply, and the page says so rather than pretending precision it doesn't have.
One more honest edge: only instrumented calls appear. A model used outside the recorder's view is absent, not shown at zero — an absence which is itself informative, because it means something is calling models off the books. Finance teams have learned to read cloud bills for shadow infrastructure; this page gives them the same instinct for shadow AI, and the reconciliation is simple — if the vendor invoice says more than the page does, something un-recorded is spending, and that's worth a conversation regardless of the dollars.
Where LLM money actually goes
The token columns are where finance and engineering finally have the same conversation, so the page explains them the way you'd want them explained:
- Input is everything sent to the model — the prompt, the carried-forward conversation, tool results fed back in. This is where surprise spend hides: history grows with every turn while the visible question stays short.
- Output is what the model wrote back. It's priced well above input at most vendors, so a modest-looking number here can dominate the bill.
- Cache read is input served from the vendor's prompt cache at a steep discount — the good column.
- Cache write is input placed into that cache at a premium. Writes far outnumbering reads means you're paying to fill a cache nobody hits — one of the cheapest optimisations you'll ever find.
Reading the table takes about a minute once you know the pattern. Scan the spend column for the model that dominates — there usually is one. Check its input-to-output ratio: an agent with enormous input and modest output is carrying conversation history it may not need. Then glance at cache write versus cache read, because a cache being filled and never hit is money evaporating on schedule. Three glances, and you know where the month went.
Every dollar shows its working
No figure on the page is unattributed: each row's cost names the rate it was priced at. When a number looks wrong, you can argue with its source instead of with the dashboard — which is the same philosophy as the rest of the product, applied to money. It also makes the page safe to forward: a screenshot that carries its own assumptions survives the finance meeting that a bare number wouldn't.
Why cost lives in an audit product
Because spend is behaviour. A runaway agent shows up here as a cost spike before the invoice arrives, right alongside the frequency anomaly it probably also triggered. "Which models are we actually using?" is an inventory question — the shadow-AI cousin of "which agents do we have?" — that governance teams need answered from evidence. And the practical ritual is monthly: five minutes with this page before the vendor invoice lands, and the invoice stops containing surprises. The recorder was already watching every call; showing you what those calls cost is the tape earning its keep twice.
Questions about how token capture works in your stack? Ask us.