Cost scenario · Automation
A code assistant in your editor
An assistant that answers questions about the codebase, explains unfamiliar functions and drafts changes. It sees the open file, related files, and the conversation.
Priced per developer per month rather than per request, because that is the number a team actually decides on.
The shape of one interaction
A 900 token system prompt, around 8,000 tokens of code context — the open file plus imports — and conversation history. Output is an explanation or a diff. An active developer averages roughly 40 interactions a day.
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 |
|---|---|---|---|
| 5 devs, one month4,400 interactions / month | $8.18$8.18 uncached | $127.60$127.60 uncached | $661.32$792.00 uncached |
| 25 devs, one month22,000 interactions / month | $40.92$40.92 uncached | $638.00$638.00 uncached | $3,306.60$3,960.00 uncached |
| 200 devs, one month176,000 interactions / month | $327.36$327.36 uncached | $5,104.00$5,104.00 uncached | $26,452.80$31,680.00 uncached |
On this shape the spread between the cheapest model that fits and the frontier one is 96.8×, which is the largest single decision on this page. Enabling prompt caching removes a further 17% on Claude Fable 5, because 30% 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
Code tokenizes densely
Source code runs around 2.7 characters per token against roughly 4 for English prose. The same screenful of code is materially more tokens than the same screenful of text, and every estimate built on the prose ratio understates it.
Context selection
Sending the open file plus every import is easy and expensive. Sending the open file plus the three symbols actually referenced is harder and much cheaper.
Interaction frequency
The variable that decides the per-seat cost is how often it gets used, and that varies by more than an order of magnitude between developers on the same team.
How to cut it, in order of return
- Retrieve relevant symbols instead of whole files40–60%
The largest lever here. Most of a file is irrelevant to any given question, and the retrieval work pays for itself immediately at this volume.
- Cache the system prompt and project conventions20–25%
Coding assistants carry long instruction blocks about style and conventions. They are perfectly stable and re-sent constantly.
- Route completions and explanations differently30–50%
Explaining a function and architecting a refactor are not the same task. The first runs fine on a mid tier; only the second needs the top one.
What people get wrong budgeting this
- Per-seat estimates built on average usage understate the bill, because usage is heavily skewed — a few developers generate most of the traffic.
- Token counts for code are underestimated by every characters-divided-by-four rule. Count real files.
- Agentic coding features multiply this several times over, because they run loops rather than single calls.
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 Token CounterReal BPE tokenization, not characters ÷ 4. Shows which counts are exact and which are estimates.
- AI Cost per User CalculatorPer-user cost with the heavy-user tail modelled, and margin against your price.
- LLM API Cost CalculatorRequests per month in, dollars out. Input, cached input and output priced separately.
- Tokenizer PlaygroundEvery token rendered separately, with its ID. The fastest way to understand why a prompt is expensive.
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.