What LLM Usage Actually Costs an Email Tool to Run

What LLM usage costs an email tool is the question that determines whether a subscription price feels fair, why cheap tiers throttle the model, and why bring-your-own-key exists at all. The short answer is that the bill is dominated by output tokens on a handful of drafts and summaries, not by the thousands of triage passes that touch every incoming message. That is why the same tool can charge one plan a fraction of a cent per action and another plan two dollars a month.

This is a decision-guide, not a marketing page. The goal is to give you enough of the underlying math that you can look at any vendor's plan — including ours — and predict whether the pricing works for your inbox. Every figure below is a ratio or a shape rather than a fixed dollar; the underlying provider prices from OpenAI, Anthropic and Google change every few months and any number pinned here would be wrong by the time you read it. We build AI Emaily, so we have skin in this — and we route AI calls through OpenRouter, which is why the model discussion below applies to us before it applies to anyone.

What one email actually costs in tokens#

Before you can reason about a monthly bill, you have to reason about a single message. Every AI action on an email has three token counts: what the model reads (input), what the vendor's own prompt scaffolding contributes (system), and what the model writes (output). Providers price input and output separately, and output is always more expensive per token — usually four to five times more, sometimes more than that on frontier reasoning models.

A rough shape for a personal inbox, verified against public API pricing as of August 2026:

Two things about this table matter more than the exact numbers. First, triage is genuinely cheap — a few hundred input tokens against a small model rounds to a fraction of a cent per email. Second, drafts are where the money goes, because 400 tokens of output on a mid-tier model costs more than 4,000 tokens of input on the same model. If a vendor tells you they can afford to run AI on everything but has to meter drafts, that is why.

The four levers that actually move the bill#

Every serious cost-reduction technique in this category is a variation on one of four levers. A tool that pulls three of them well is often ten times cheaper to run than one that pulls none — which is the whole reason vendor prices for what looks like the same feature can differ so much.

Model right-sizing is by far the biggest of the four. A small classification model costs roughly one to two orders of magnitude less per token than a frontier chat model, and a triage decision does not need a frontier model to be correct. A vendor that runs the same model for triage and drafting is either overpaying for triage or underpowering drafts; a vendor that uses a gateway to route each task to a fit-for-purpose model can afford both.

This is the argument for routing through a gateway rather than a single provider. AI Emaily uses OpenRouter, which sits in front of dozens of models from OpenAI, Anthropic, Google, Mistral, Meta and others. The routing decision — small model here, mid-tier there, frontier only when the task warrants — happens per action, not per plan, so a busy triage stream does not push you into a bill it did not need to generate.

A scoring rubric for reading any vendor's cost story#

When you look at a vendor's pricing page or their AI credit table, the question is not what the number is — the number will be out of date within a quarter. The question is whether the underlying architecture can sustain that number without either throttling you or losing money on your seat. Here is the rubric.

A worked example: one moderate inbox, one month#

The numbers below use published OpenAI and Anthropic API list prices as of August 2026 and are illustrative — verify against the OpenRouter or provider pricing page linked at the bottom before quoting them. The point of the exercise is the shape of the answer, not the last cent.

Assume a moderate professional inbox: 60 incoming emails per working day, 20 working days, so 1,200 messages per month. The user runs triage on every message, drafts a Copilot reply for 10 messages per day, summarizes 30 threads across the month, and asks 20 questions of their inbox. The user is on a mid-tier draft model (Sonnet-class or GPT-4-class) with a small model for triage (Haiku-class or GPT-4o-mini-class).

1. 1

   Triage — 1,200 messages

   Around 300 input and 40 output tokens each on a small model. Even at list price and without caching, this whole layer costs less than the price of a single frontier-model draft. Triage is the loud part of the product and the quiet part of the bill.

2. 2

   Copilot drafts — 200 per month

   Around 2,500 input and 400 output tokens each on a mid-tier model. Prompt caching pulls the input cost down substantially, because the same context brain and recent history are reused across a session. Output — the actual reply — is what you pay for.

3. 3

   Thread summaries — 30 per month

   Around 2,000 input and 100 output tokens each, mid-tier model. Cheap in absolute terms; summaries are short by design.

4. 4

   Ask over the inbox — 20 per month

   Around 5,000 input tokens with retrieval, 200 output tokens each. The retrieval step matters — a well-scoped retriever keeps this cheap; a naive one that dumps every match into context makes Ask the single most expensive line on the bill.

Add the four lines together and the total model bill for that inbox typically sits in the low single digits of dollars per user per month — before the vendor's own infrastructure, storage, sync, support, and margin. That is the honest reason a $15 to $30 per-seat plan is not a rip-off: the model bill is a fraction of the price, but it is not zero, and it is the fraction that scales with how hard you use the AI. It is also why a $5 plan that promises unlimited drafts on a frontier model cannot really mean it. Either the model gets downgraded silently, the rate limit kicks in early, or the tier is a loss-leader that will not survive the next pricing round.

Shift the same inbox to Autopilot with autonomous sends and the bill roughly doubles on the draft line, because each send is two runs — the draft and the send — and the send draft is usually not the same short reply a human would type. Shift again to a high-volume inbox at 200 messages per day and the triage layer alone starts to matter, though it is still smaller than the draft layer. Volume changes the composition of the bill more than it changes the total.

Red flags in vendor cost stories#

Once you know the shape of the math, some vendor claims stop being plausible. These are the four that come up most often, and the reason each one should make you look closer.

• Unlimited frontier-model drafts on a plan cheaper than what a moderate inbox would cost the vendor at provider list price. Either the tier is a promotion, the model is being swapped, or there is a rate limit that has not been documented.
• No mention of which models run. If the vendor cannot tell you whether triage runs on a small model and drafts on a bigger one, the answer is usually that everything runs on whichever model was cheapest to integrate — and one of the two workloads is being overpaid for.
• Silence on caching. Prompt caching is table stakes as of 2026; a vendor that has not implemented it is paying for input tokens they should be reusing, and that cost is being paid out of your seat.
• No BYOK path. Bundled compute is fine as a default and often cheaper for light users, but a vendor with no bring-your-own-key option is telling you they need the markup on the model bill. That is a legitimate business choice; it is a warning sign only when combined with unlimited claims.

What we think, and where it does not apply#

We build AI Emaily. Because we route through OpenRouter, we can right-size the model per task — a small model on triage, mid-tier on drafts and summaries, frontier only when the reasoning warrants it — and take advantage of prompt caching on the parts of the prompt that repeat, like your Personal Context brain and the recent thread history. That is why our pricing is metered in credits per action rather than per token: the token math sits underneath and the credit price stays predictable on your side. Triage is cheap enough to run on every incoming message; drafts are the metered action, which is the honest place to meter. BYOK is available for users who would rather pay a provider directly.

Where this does not apply, and where an honest post has to say so: if you are a heavy power user who processes thousands of messages a day and you already have a negotiated enterprise rate with a single provider, a pure BYOK-only tool that adds no compute markup will be cheaper on the pure model bill than any vendor with bundled credits, ours included. The bundled pool exists so most users do not have to think about the model bill; if you want to think about it, the shape of the answer is different, and a client that lets you plug your key straight in and skip the pool is the right fit. On that specific dimension, some competitors that ship pure BYOK-only with no bundled compute at all give you a tighter cost floor than a bundled plan can, and it is worth being upfront about that. AI Emaily has a 7-day free trial on Pro and Autopilot — a card is taken and nothing charges if you cancel before day seven, which is enough time to see your own credit usage on a real week of mail. See the AI Emaily homepage for the product, and the pricing page for current plan sizes.