What LLM Inference Actually Costs Per Email

How much does LLM inference cost per email has a real answer, and it is smaller than most buyers guess and larger than most vendors imply. As of August 2026, triaging one message costs a fraction of a cent. Drafting a reply costs a few times that.

The interesting part is not the headline number. It is that the same email, on the same model, with the same prompt, can cost three times more or less depending on a variable almost no comparison mentions: whether your mail arrives in clusters or one message at a time.

This post walks a single email from raw text to a line on a provider invoice, so you finish with arithmetic you can redo yourself when list prices move — which they do.

The short answer: what one email costs#

Below is a worked cost for two jobs — classifying one incoming message, and writing one reply to it — across three model tiers, with and without a warm prompt cache. The token budgets behind these figures are broken out in the next two sections, so you can substitute your own.

The rates used are round rungs, not quotes: 1, 3 and 10 dollars per million input tokens, with output at five times input. Provider list prices cluster near those rungs, but take the current figure from the provider page before you plan against it.

How many tokens is an average email?#

There is no single number, but there is a reliable conversion. Anthropic publishes the working estimate that one token is about four characters, or about 0.75 words, in English. Multiply your word count by roughly 1.3 and you are close enough to budget with.

That puts almost every real email in a narrow band. The body of a message is rarely the expensive part of the request, which is the whole point of the section after this one.

Where the tokens actually go#

An agent does not send the model an email. It sends a request that contains instructions, the actions it is allowed to take, what it knows about you, the thread so far, and then the email. The message is usually under five percent of the input.

Here is a realistic budget for the two jobs. Yours will differ, but the shape will not: a large fixed block that repeats on every request, and a small variable block that is the actual mail.

The four things that decide your bill#

Once you have the budget above, four variables move the number. They are worth understanding in this order, because the first one is bigger than the other three and gets discussed the least.

• Cache hit rate. The 2,500-token repeating block can be served from cache at a tenth of the input rate — but only if a previous request wrote it and the cache has not expired.
• Model tier. Across the published ladder, base input rates span roughly an order of magnitude. Triage and drafting do not need the same tier.
• The input-to-output split. Email is a long-input, short-output workload. Output is priced several times higher per token, but there is so little of it that the output rate barely moves the total.
• Whether the work can wait. Both major providers publish an asynchronous batch tier at half the standard rate on input and output alike.

Which cost lever is worth pulling#

Not all savings are the same size, and the cheapest ones to implement are usually the smallest. This is the honest ordering.

Worked example: one inbox for a month#

Take 1,000 incoming messages. Triage all of them; draft a reply to 15 percent. The strategy you pick moves the total by more than a factor of ten on identical mail.

What that means per user, per month#

Now scale it by inbox volume. This is the table that explains most of the pricing behaviour you see in this category.

Why AI email tools cap usage#

Read the bottom-right cell again. A heavy user running frontier inference with no caching consumes several times a typical seat price in raw tokens before anyone has paid for engineering, storage, sync or support.

That is the entire explanation for usage caps, credit systems and tier-gated AI. A flat seat price is an average bet, and email volume is not distributed evenly — a small share of users generate a large share of messages. Vendors either cap, meter, route to cheaper models, or quietly lose money on their best customers.

The structural moves in this market follow the same logic. Google folded Gemini into Workspace business plans with a per-seat rise; Microsoft made Business Standard and Premium with Copilot permanent SKUs after its July 2026 price change; Superhuman moved AI into a higher business tier post-acquisition; and the helpdesk tier went metered per resolution.

None of that is a pricing trick. It is what happens when a variable cost sits underneath a fixed price.

Red flags when someone quotes you a cost per email#

Most published per-email figures are technically true and practically useless, because they omit the variable that moved them.

• A per-email price with no model tier named. The same prompt spans a tenfold range across the published ladder.
• A per-email price that excludes thread history. Triaging a fresh message and triaging the ninth reply in a thread are not the same request.
• Unlimited AI with no published cap. There is always a fair-use clause; ask to see it in writing before you plan volume around it.
• Credits with no token disclosure. A credit that does not say how many tokens it buys is a unit of account, not a unit of measure. Ask what one triage and one draft consume.
• No answer on whether prompt caching is enabled. It is the biggest single lever and a vendor that has not implemented it is paying two to three times more than it needs to — which eventually reaches you.

What we would pick, and who should pick something else#

Start with the concession, because for a lot of readers it ends the decision. If you already pay for Microsoft 365 or Google Workspace on a tier where Copilot or Gemini is bundled, your marginal token cost for basic drafting help is zero. Nothing in the arithmetic above beats a cost you have already paid. If drafting assistance inside the mail app you already use is genuinely all you want, stop here and use what you own — the comparison worth having is assistant versus agent, and it is a capability question, not a cost one.

Second scope note: on pure token cost, a self-built pipeline wins. Small-tier triage at batch rates with a well-ordered cache prefix undercuts every subscription here. What a subscription buys is the sync engine, the approval surface, the undo path and the audit trail — not the tokens.

We build AI Emaily, so weigh the next paragraph accordingly. Our answer to this problem is the routing split the table above argues for: calls go through a gateway that picks a model by task and tier, so triage does not pay frontier rates. Usage on our keys is metered against plan caps, and bring-your-own-key is supported, with keys decrypted only in an isolated worker and never client-side or logged — which means if you want to see the raw provider invoice for your own inbox, you can. There is no permanent free tier; it is a 7-day free trial on Pro and Autopilot, card required.

We are the right pick if you want the agent to act — triage, file, draft, follow up — with approve-before-send, undo and an audit trail, across Gmail, Outlook and IMAP in one place. We are the wrong pick if your only question is token cost and you have the engineering time to build the pipeline yourself; if you need a Linux build, which we do not ship; or if you want a native-toolkit Mac client, since our desktop app is an Electron shell around the web interface and will not match a native binary on memory footprint or deep OS integration.