Drafts are written by a language model — a program that understands text and can write it. draften calls it over the internet at a provider (Anthropic by default; OpenAI or a model on your own server work too), and every use is paid for. Here is roughly what it costs and how you keep it under control.
Three tasks, three models
The assistant does not use the same model for everything. A cheaper one does the simple work; the strong one only writes:
| Task | What it does | Default model |
|---|---|---|
| Triage | decides which new mail is the assistant's | Claude Haiku 5.5 (cheap) |
| Drafting | writes the draft, reading the knowledge sources | Claude Sonnet 5.5 (strong) |
| Learning | evaluates your edits and proposes lessons | Claude Haiku 5.5 (cheap) |
Which model and key the assistant uses is in Settings → Models and keys. You usually need not change them.
Assistants that wrote with Claude Sonnet 5 were switched to Sonnet 5.5 at the same price with the draften update, and triage and learning from Claude Haiku 4.5 to Haiku 5.5 at a tenth of the price; assistants on older OpenAI GPT-5 models were switched to GPT-6 models. You see the new model in the assistant's Settings and on every new draft; the switch does not appear in the change history.
Model effort
Effort is how thoroughly the model thinks before it answers: a higher one tends to be more careful, but takes longer and costs more. It is set for each task separately — in Settings → Models and keys, the Key and model per task card, the Model effort field under the model in the Triage, Drafting and Learning boxes.
The first choice, the model's default, lets the model work at its default level (the brackets show which one that is). The effort returns to it after a change of model too. Each model takes different levels:
| Model | Levels | Default |
|---|---|---|
| Claude Sonnet 5.5, Sonnet 5, Opus 5, Fable 5.1 | low to maximum | high |
| Claude Opus 5.5 | low to maximum | medium |
| Claude Haiku 5.5 | low to maximum | medium |
| Claude Haiku 4.5 | the model takes no effort setting | — |
| GPT-6 Luna, GPT-6 Sol, the GPT-5.6 series | none, low to maximum | medium |
| GPT-6 Astra | low to maximum | medium |
| a model on your own server | any | set by the server |
Before, the assistant had one effort for all tasks. It was carried over to Drafting; Triage and Learning run at the model's default level.
What it costs
Prices are in US dollars, as the model provider charges them; draften does not convert them. Roughly (a draft is 40–120 thousand input tokens, half of them from the cache, and 1–2 thousand output tokens):
| What | Per item | Typical month |
|---|---|---|
| triage of new mail (Haiku 5.5) | a fraction of a cent per check | under 1 USD |
| a draft (Sonnet 5.5 or Sonnet 5) | 0.05–0.15 USD | 5–20 USD at 3–5 drafts a day |
| a draft (Opus 5.5) | 0.10–0.29 USD | 10–40 USD |
| a draft (Opus 5) | 0.14–0.38 USD | 15–60 USD |
Long threads and drafts for which the assistant reads a lot in the sources cost more. The new-assistant wizard is a one-off cost — before every paid step it shows an estimate, usually a few dollars.
Where you see the spend
- Overview, the What the models cost card — what writing, triage and learning cost over the chosen period, today's spend and the daily cap.
- Mail and Drafts — the cost of every message and draft.
- Settings → Activity — mail checks and requests with their outcome.
Daily cap
In Settings → Behaviour and limits you set a daily cost cap. Once the assistant reaches it, it writes no more drafts that day, evaluates nothing and starts no paid wizard step or trial; work under way finishes. In the morning (by the organisation's time zone) the count starts over. Mail arriving meanwhile gets the limit status.
The cap is a good safeguard in case an unexpected flood of mail arrives.
Who pays
The model provider is paid through an API key — the organisation's, the mailbox's, or the draften operator's default key. Who manages the keys is described in the administrators' page Model API keys.