Everything about what R3 Coach AI is, what it will and will not do, how usage is measured, exactly what it costs, and how to stop it.
Last updated 2026-09-24
R3 Coach AI is an assistant that lives inside your R3 dashboard. Its job is narrow on purpose: it helps you apply the R3 products you already own, on the particular day you are actually having.
That distinction matters more than it sounds. A general chatbot will answer any postpartum question you ask it, confidently, from whatever it absorbed from the open internet. R3 Coach will not. It answers from your R3 material, and when your question falls outside that material it says so rather than inventing a plausible paragraph. That is a deliberate restriction, and it is the whole reason the answers are worth acting on.
It is not one of the R3 Continuity gifts, and it is not included in the price of any product. It is a separate capability with its own separate billing, which is why this page exists.
You do not have to use it. Everything in every R3 product is complete without it: the book stands alone, the calendar stands alone, the recipes and the workouts stand alone. R3 Coach exists for the moments when you have the plan in front of you and a specific obstacle in the way, and you would rather ask than go looking.
The requests it handles best are the small, concrete, in-the-moment ones.
The pattern is that it is at its best when the answer already exists somewhere in what you own and the work is finding it, adapting it, and handing it to you in a form you can act on in the next four minutes.
It will not diagnose anything, ever. It will not tell you whether you have an abdominal separation, a prolapse, a hernia, an infection, or a postnatal mood disorder. It can explain what those words mean. It cannot tell you whether one of them applies to you, and it will say so rather than guess.
It will not prescribe a calorie deficit, and if you are breastfeeding it will not propose a deficit at all. It will point you at eating enough and at the breastfeeding track. This is not caution for its own sake: aggressive restriction while feeding is a bad idea, and a chatbot that agrees to it because you asked nicely is a chatbot that has failed you.
It will not tell you that a change in your milk supply is safe. Supply belongs with your midwife, health visitor, lactation consultant or doctor, every single time, and there is no phrasing of the question that changes that answer.
It will not override an instruction your own clinician has given you. If your physiotherapist told you not to do something and R3 suggests it, your physiotherapist wins, and R3 Coach will tell you so.
It will not promise you a result. Not kilograms, not centimetres, not a timeframe, not a flat stomach, not closed separation, not a preserved supply. Nobody can honestly promise those, so it does not.
And it has no ability to change your account. It cannot grant you a product, change a price, waive a charge, alter your eligibility, or lift any limit. Those are decided by the system, by code, and R3 Coach is not given a lever for any of them. If you ask it to, it will tell you plainly that it cannot, because saying anything else would be untrue.
Before every answer, the system pulls the relevant passages out of the R3 products on your account and hands only those to the model. It is not browsing, it is not searching the web, and it is not drawing on a general pool of postpartum advice.
That has a consequence worth stating plainly: it cannot quote a product you have not bought. If you ask about cesarean progressions and you do not own The C-Section Return Guide, it will tell you the answer lives in that guide and name it. It will not paraphrase it from memory. This is not an upsell tactic, it is the same entitlement rule that governs your downloads, applied to the assistant.
It also carries a small amount of durable context about you, if you give it: roughly how far postpartum you are, whether you are feeding, what equipment you have, how much time you usually get, your household size, and anything you have told it to avoid. That is the complete list. It is kept separately from your conversations, and it is editable and deletable from your dashboard.
Everything else is the recent conversation, which is trimmed as it grows. A conversation that grows without bound is a bill that grows without bound, so only the last several turns travel with each message.
It is worth understanding, because it explains both the quality of the answers and the shape of the bill.
Every R3 product is stored not as one long file but as a set of small labelled passages: each meal, each core drill, each session, each plateau scenario, each safety rule. Every passage carries the id of the product it belongs to.
When you send a message, the system scores those passages against what you asked, filters them to the products on your account, and takes the best handful. The safety rules and the description of the method travel with every single answer regardless of what you asked, because an assistant that forgets its limits when the question is interesting is not safe.
Those passages are then handed to the model inside an explicit data fence, labelled as content rather than as instructions. That labelling is not decoration: it is what stops a recipe, a note you saved, or something you pasted in from being able to issue orders to the assistant. If any of it says "ignore your rules", the assistant treats that as text to discuss, not a command to obey.
The practical consequence for your bill is that input is usually the larger side of a message, and that caching does most of the work of keeping it cheap.
R3, including R3 Coach, provides general educational information about nutrition, movement and strength after birth. It is not medical advice, it is not diagnosis, it is not treatment, and it does not replace care from your own clinician.
Some questions never reach the model at all. If your message mentions bleeding that is getting heavier, pain that is sharp or worsening, an incision that is red, hot, open or discharging, a visible bulge or doming along the midline under effort, leaking or a feeling of heaviness or something coming down, a fever, a change in your baby's weight gain, or anything about not coping, the system short-circuits and returns the conservative answer directly.
That answer is always shaped the same way: here is what this needs, here is who to talk to, and here is what R3 is until then, which is walking and breathing work and nothing else. It is identical every time, because it is the right answer every time, and it is not generated by a model that might phrase it differently on a bad day.
A short-circuited safety answer costs you nothing. It is returned before any tokens are spent, so you are never billed for being told to ring your midwife. That is deliberate: a worried customer hesitating because she is watching a balance would be a genuinely bad outcome.
If you are in immediate danger, or you are worried about your own safety or your baby's, contact your local emergency services. Do not ask a chatbot.
This is the section most people come here for, so it is stated as plainly as we can manage.
Language models do not read characters or words. They read tokens, which are chunks of text roughly three quarters of a word long on average in English. A short question might be twelve tokens. A paragraph of reply might be ninety. The R3 material handed to the model alongside your question is usually the largest part, and it is counted too.
Every message therefore has two sides: input, which is everything sent to the model, and output, which is everything it writes back. They are priced differently because they cost differently.
Your price is derived from what the underlying AI provider charges us, multiplied by 20. That multiple covers the provider bill, the infrastructure around it, the retrieval layer that keeps the answers grounded in your own material, and the safety machinery described above. It is a fixed multiple, published here, and it does not change per customer.
| What is counted | Per 1,000 tokens | Per 1,000,000 | When it applies |
|---|---|---|---|
| Input tokens | $0.0400 | $40.00 | Everything sent to the model: your question, the recent conversation, and the R3 passages it was given. |
| Output tokens | $0.2000 | $200.00 | Everything the model writes back to you. |
| Cached input | $0.004000 | $4.00 | Input served from cache, which is most of the R3 material on every message after the first. |
| Cache write | $0.0500 | $50.00 | The first time a block of content is written into the cache. |
The rates above are the current customer-facing rates in US dollars. They are generated on this page from the live configuration, so what you are reading is what you would be charged right now, not a number typed into a document months ago.
After every answer, the AI provider reports exactly how many input tokens, output tokens, cached-input tokens and cache-write tokens that request consumed. We take those numbers, not an estimate. We do not count characters, we do not approximate from message length, and we never accept a token count sent by your browser, because a number that arrives from a browser is a number that can be edited.
Each category is priced at its own rate from the table above, added together, and written to your usage ledger as one row. The row records the model, the timestamp, the four token counts, the provider cost, the multiplier applied, and the resulting charge. The rate in force at the time is stored on the row itself, which means that if rates ever change, your history is not retroactively repriced. You can see every row in your dashboard under Billing.
Usage accumulates as an unsettled balance rather than being charged message by message. Nobody wants fourteen card transactions of eleven cents each, and no card processor wants them either.
Settlement is idempotent, which is a technical way of saying it cannot fire twice for the same balance. Each settlement carries a key built from your account, the billing week, and the exact amount. A retry after a network timeout finds that key already used and does nothing.
It is easier to trust a number you can follow, so here is one, start to finish.
Suppose you open a new conversation and ask: "I slept about four hours, the baby is teething, what should I actually do today?"
Now ask a follow-up in the same conversation: "make it five minutes instead." This time most of that 2,000-token block of R3 material is identical to last time, so it is served from cache at the much lower cached rate, and only the new part of your message is charged at the full input rate. The second message therefore costs a fraction of the first. This is the single biggest reason a long conversation is cheaper per message than a series of one-off questions, and it is why we do not charge a flat per-message fee: a flat fee would quietly overcharge exactly the people who use it most sensibly.
We could have published a flat price per message. We did not, for two reasons.
First, messages are not the same size. A one-line question about a snack and a long back-and-forth about restructuring your week cost very different amounts to answer, and a flat fee means the short question subsidises the long one. Charging for what you actually used is the fairer arrangement, even though it is the harder one to explain.
Second, a published multiple is auditable in a way a flat fee is not. The provider publishes its rates. We publish the multiple. You can therefore check our arithmetic, which is not true of a number we simply assert.
The multiple covers more than the provider invoice. It covers the retrieval layer that pulls your own material in front of every question, the entitlement checks that keep it to what you own, the safety layer that intercepts anything clinical before it reaches a model, the ledger that records every row, the settlement machinery, the failed-payment handling, and the support time when something goes wrong. Those are the parts that make the difference between a chatbot and something worth putting next to a health product.
If you also have R3 Continuity and R3 Daily Autopilot running, they stay on separate lines with separate descriptions. A week might read:
| Line | Amount |
|---|---|
| R3 Continuity | $25.00 |
| R3 Coach AI usage | $12.40 |
| R3 Daily Autopilot, week of 6 April | $17.50 |
They are never collapsed into one unexplained total. If you cannot tell what a charge was for by reading its description, that is a bug and we want to hear about it.
An average question and answer is a few hundred input tokens plus the R3 passages, and one or two hundred output tokens. In practice most questions land in the region of a few cents each once caching is doing its work.
Caching matters here, and it works in your favour. The R3 rules and much of the retrieved material are identical from one message to the next, so after the first message in a conversation most of that input is served from cache at the much lower cached rate in the table above. A long conversation is therefore considerably cheaper per message than its first message suggests.
Someone who asks a handful of questions a week will usually sit under the $7 minimum for a while and see nothing charged at all for the first stretch. Someone leaning on it daily, in long conversations, will see a settlement most weeks. Your actual number is on your dashboard, updated after every message, and you never have to work it out from this page.
There is one more protection: a daily ceiling of $20 of AI usage per account. Reaching it stops further billable use until the next day. It exists so that a runaway loop, a forgotten open tab, or somebody else getting into your account cannot produce a bill you did not choose.
R3 Coach AI is only available when there is a payment method on file that can legitimately be charged again. In practice that means a card, or a PayPal wallet you authorised for future payments at checkout.
If you paid with Apple Pay or Google Pay, there is a genuine technical reason it is not available: both wallets give merchants a single-use device number rather than a reusable card. There is nothing on file that can be charged a second time. That is a privacy feature of those wallets working exactly as designed, and it means anything billed by usage cannot be attached to them.
For those accounts, R3 Coach is not hidden behind a message: it is genuinely off. The navigation does not show it, the route refuses server-side, and no usage can be recorded. Everything you bought still works in full.
If you want access later, you can add a card deliberately from Billing in your dashboard. That is an explicit action you take, never something that happens because you clicked past a screen.
Before the first message that could ever be billed, you see a consent screen. It states that AI usage is billed separately, that usage varies with how much you use it, that the payment method you already have on file may be charged, what the current rates are, when settlement happens, where to see your usage, and how to stop.
You tick a box. Nothing is pre-ticked. If you do not tick it, nothing happens and nothing is billed.
After that, we do not interrupt you again. A product that asks you to reconfirm every time you type is not being careful, it is being annoying, and it teaches you to click past warnings. You agreed once, knowingly, and the running balance is visible whenever you want it.
Under the chat box there is a permanent line: "AI Chatbot. Extra usage may apply. Learn more". The "Learn more" is the link that brought you here. It is on every screen where you can type to R3 Coach, and it never disappears.
Your conversations are stored against your account so you can reopen them. They are readable by you. They are not shown to other customers, and the system enforces that in two independent ways: conversations are keyed by your account, and every read re-checks that the stored record belongs to the account asking for it. One check would probably be enough. Two is what a cross-account leak deserves as a precaution.
Durable memory is deliberately small and deliberately named. Only these fields are kept: your first name, roughly how far postpartum you are, whether you are feeding, your equipment, your usual available time, your household size, foods you avoid, and allergies. Nothing else from a conversation is promoted into durable memory. A system that quietly accumulates everything you ever mentioned about your body is a liability you never agreed to, so it does not do that.
You can view, edit and clear that profile from your dashboard, and you can delete any conversation. Deleting a conversation removes it.
Your health information is never sent to advertising platforms. Not your feeding status, not your delivery type, not symptoms, not weight, not measurements, not anything you typed to R3 Coach, not anything you configured in Autopilot. Advertising events carry a purchase value, a currency and a hashed identifier, and that is the entire list. This is written down in more detail in our privacy notice.
Your messages are processed by a third-party AI provider in order to produce an answer. They are sent for that purpose and no other.
Several limits run quietly in the background. They exist to stop a compromised account, a runaway script or an honest mistake from producing a bill nobody intended.
None of these affect anything you have bought. A paused assistant never removes access to a PDF, a recipe, a workout or your membership content.
Cards expire and banks decline things. If a settlement fails, the balance stays on your account as outstanding and we try again in line with what the payment provider supports.
After repeated failures, billable AI use is paused. That is the honest option: the alternative is letting an unpaid balance grow on an account we already cannot charge, which serves nobody.
Your dashboard will say plainly what happened and what the outstanding amount is, and you can update your payment method there. When it goes through, AI use resumes.
Throughout all of it, everything you already bought stays yours. A failed AI settlement never removes access to a product, a download or your membership. Those are separate things and they are treated as separate things.
Stopping is simple, and there is no retention flow to click through.
R3 Coach is built on a large language model, and large language models get things wrong. It is grounded in your own R3 material specifically to reduce that, and it is instructed to say when something is outside what it was given rather than fill the gap. It will still occasionally be wrong, and the R3 product itself is the authority whenever the two disagree.
It is not a clinician, it is not a physiotherapist, it is not a dietitian, and it is not a lactation consultant. For anything to do with your body, your recovery or your baby, those people are the answer and R3 Coach will tell you so.
Availability depends on a third-party provider. It can be slow or briefly unavailable. Nothing else in R3 depends on it: your products, your calendar, your recipes and your workouts all work whether or not the assistant is up.
If a charge does not look right, or something on this page does not match what you are seeing in your dashboard, write to us at support@r3-method.com. Include the date and the amount and we will find it.
We would genuinely rather refund a charge you did not understand than keep it.