Most conversations about AI costs are about tokens. That's right when you run a frontier model: the token bill is nearly the whole bill. But model prices now span about 100x from the top tier to the small models, and when tokens get cheap the bill doesn't go away. The rest of the stack becomes the bill.
To see where that leaves a finance team, we modeled one AI product at three model price points and held everything else constant.
The model#
One product serving 1M agent requests a month. Each request reads 20k input tokens (instructions, retrieved documents and every step of the agent loop), 40% of them from cache, and writes 1k output tokens. Around the model sit the usual lines: compute and retrieval, logs and monitoring, storage and egress. Only the model price changes between the three bars.
Move the sliders to put in your own numbers.
On a frontier model, tokens are 92% of the bill. On a small model they are 11%, and compute + retrieval, egress and storage move up the list.
| Line | Frontier | Mid | Small |
|---|---|---|---|
| Model tokens | $172k · 92% | $35k · 71% | $1.8k · 11% |
| Logs + monitoring | $1.4k · <1% | $1.4k · 2.9% | $1.4k · 8.8% |
| Compute + retrieval | $8.0k · 4.3% | $8.0k · 16% | $8.0k · 49% |
| Storage | $2.3k · 1.2% | $2.3k · 4.7% | $2.3k · 14% |
| Egress | $2.7k · 1.4% | $2.7k · 5.5% | $2.7k · 17% |
| Total | $186k | $49k | $16k |
What changes#
- Frontier model ($10 / $50 per million tokens): about $186k a month, 92% of it tokens. The first cut is obvious: route easy requests to a cheaper model, cache more, and send less context.
- Mid model ($2 / $10): about $49k a month, 71% tokens. Tokens still lead, but compute and retrieval are now 16% of the bill.
- Small model ($0.10 / $0.50): about $16k a month, 11% tokens. Compute and retrieval are 49%, egress 17% and storage 14%. The model is no longer what finance asks about.
What surprised us#
We expected logs to rise to the top on the cheap model, because agents write a lot of them. At list prices they don't. Twenty indexed log events per request is 20M events a month, about $34 to index. Monitoring only becomes a large line when the volume gets very large, or when it's priced per host across a big fleet.
What does rise is everything that touches the data: the compute that finds the right documents for each request, the data that crosses network boundaries to reach the model, and the data you keep. Once tokens are cheap, the bill is driven by where your data sits, how much of it moves, and how much work it takes to find the piece a request needs. That's the part of the stack we build at Miriel.
Caveats#
Your mix will differ. The biggest assumption in this model is $8 of compute and retrieval per 1,000 requests, so try it at your own figure. Prices are list prices as of October 2026: model APIs from Anthropic's pricing page, logs and monitoring from Datadog's list prices, and storage and transfer from Amazon S3 pricing. Committed-use discounts lower every line by different amounts, so check the order against your own bill.
