Estimate tokens before sending (usage preview and confirmation gate)

Large notes, long conversations, or many attachments can use a lot of tokens (cost) in a single send. This feature estimates that before you send and, if needed, asks you to confirm (off by default).

Enable it: Settings (⚙) → AI → "Estimate tokens before sending". When off, nothing is computed (no badge, no confirmation — zero friction, as before).

Live badge: when on, the estimated input tokens appear to the left of the send button (e.g. "~1.2k") and update as you type. This is an approximation on the same scale as the pre-send confirmation: it includes the fixed overhead (instructions, tool definitions, etc.). Right after opening it may briefly show a smaller number until that overhead arrives.

Confirmation gate: on send, if the estimate exceeds your threshold, a prompt appears — "This request will use about N tokens. Send it?" (with "Estimated cost: ~$X" below when the price is known) — with Send / Cancel. Below the threshold it sends without asking (no friction for small requests).

Resend and queued messages too: the error card's "Resend" and messages queued while a reply is generating go through the same check right before they are sent. If the estimate itself fails, you'll see "Couldn't estimate the tokens" once and the message is sent without confirmation (so a broken estimate never blocks sending).

• Estimate vs. actual: the estimate counts only the input of the first request and does not include prompt-caching discounts or surcharges. When the AI goes back and forth (for example, reading notes), the ↑ under the reply is the total of all round trips, so it can be larger than the estimate.

How attachments are counted: text files and PDFs are counted in the form they will actually be sent (the embedded text, the PDF's page count, and so on). When a PDF goes to Anthropic as a document, that PDF alone is sent once to Anthropic's free token counter after you press send (before the confirmation) so the number is accurate. The estimate shown under the input box is calculated locally and sends nothing. If you cancel at the confirmation, that one count has already been sent (the same file isn't sent again after the first time).

Change the threshold: in the same settings spot, set "Confirmation threshold (tokens)" (default 15000). Lower asks more often; higher asks less.

Cost estimate: when the selected model’s price is known, "~$X" is also shown (input tokens on your own API key; output is not included because it is not known before sending). For models with unknown pricing, only the token count is shown.

Actual tokens consumed after sending are shown below each reply (see "Using chat").