Skip to main content
This page is the commercial view of the plan catalog — who each tier is for, what a conversation actually costs, and what sales should and should not promise. For the mechanics of credits see Credits and Usage; for the in-product upgrade flow see Plans and Pricing.Every figure below is measured, not modelled — the source is a replay of 132,353 real billing events on the development platform (2026-09-06). Where a number is an estimate or unverified, it says so.

​
The catalog

The per-credit rate falls as the tier rises — that is the volume discount, and it is enforced at the database level: a plan that sold credits at a higher unit rate than a cheaper plan would fail the pricing migration outright.
Scale is priced but switched off. The catalog row is coherent and the product now supports selling it — the upgrade page renders any active plan and the checkout flow accepts any plan that resolves to a Stripe Price, so it is no longer a code change. Two things are still outstanding: a Stripe Price at 499and‘isactive‘ontherow.Untilanadmindoesboth,sellScaleasa∗∗sales−assisted∗∗tierandprovisionitmanually,exactlylikeEnterprise.Donotputa"buynow"buttonnextto499 and `is_active` on the row. Until an admin does both, sell Scale as a **sales-assisted** tier and provision it manually, exactly like Enterprise. Do not put a "buy now" button next to 499 before checking the page actually shows one.

​
Credit top-ups

When a customer runs out mid-month they can buy credits without changing plan. Add-on credits do not expire and are spent only after the month’s plan allowance is gone. A top-up is always dearer per credit than any plan, and that is deliberate. If overflow were cheaper than upgrading, a customer at the Growth ceiling would buy top-ups forever instead of moving to Scale, and we would sell the same credits for less at the same cost. The floor is the cheapest paid plan’s rate (Starter, $0.004875) and it is enforced in the database: a catalog where a top-up undercuts a plan fails the pricing migration outright. So the upgrade is the better deal at volume, and you can say so with numbers. A Growth customer who needs another 95,000 credits pays 350moretomovetoScale—350 more to move to Scale — 0.0037 per extra credit — against $0.0050 buying the largest top-up. Upgrading is about 26% cheaper. Position top-ups as what they are: a bridge to the end of the month, not a way to buy volume.

​
Changing a plan’s price

Three steps, in this order, and none of them is a deploy:
  1. Create a new Price in Stripe on that plan’s Product, at the new amount. Never edit an existing Price — Stripe does not allow it, and it is the one operation that would move customers who are already paying.
  2. Point the plan at it. platform_billing_plans.stripe_price_id is the source of truth for which Price a checkout session uses; setting it takes effect within a minute, with no redeploy.
  3. Change the advertised numbers (monthly_price_usd, monthly_credits).
Existing subscribers keep their old price AND their old allowance. Stripe pins each subscription to the Price it was created with, so changing what we SELL never changes what a current customer PAYS. Their credits and limits are pinned too: every account carries a frozen copy of the terms it signed up on, so a repriced plan does not change what an existing subscriber is shown or granted at their next renewal.That is a promise you can make in a sales conversation: your price and your allowance are fixed for as long as you stay on this plan. Moving existing customers onto new terms is a separate, deliberate admin action — never a side effect of repricing — and the admin Plans page shows how many subscribers are on older terms before anyone takes it.Every deploy checks this: a plan advertising an amount its bound Stripe Price does not charge fails the deploy rather than going live, so the upgrade page and the customer’s card can never quietly disagree.

​
Who each tier is for

Free — the evaluation

Someone deciding whether an AI agent can answer their questions at all.1,000 credits is roughly 14 conversations. That is enough to wire one agent to one knowledge base, ask it the ten questions that matter, and see the answers. It is not enough to run a pilot, and it is deliberately not enough to ingest a large website.Headline: Try it on your own content. No card.

Starter — $39

One person, one agent, one job. A solo founder on a docs bot, a consultant putting an assistant on their own material, a team running a single internal FAQ agent.8,000 credits ≈ 116 conversations a month. (A typical conversation is cheaper than that implies — see “What a credit buys” — but a month’s allowance is spent by a mix, and the mix is what the allowance divides by.) Three agents and two knowledge bases.Headline: A production support agent for less than an hour of your time.

Growth — $149

A team with real, recurring inbound. Customer support, sales qualification, or an internal helpdesk that people actually use daily.35,000 credits ≈ 507 conversations a month. Ten agents, five knowledge bases, fifteen seats — enough to run different agents for different audiences instead of one compromise agent.Headline: Answer 500 customers a month for less than one support hire’s day rate.

Scale — $499

High, steady volume, and knowledge bases that keep growing. Deflecting a meaningful share of a real support queue, or running agents across several product lines.130,000 credits ≈ 1,885 conversations a month, and the headroom to re-ingest large sources without watching the meter. Unlimited agents, knowledge bases and seats.Headline: The cheapest per-answer rate we sell.
Enterprise is a contract, not a tier: custom credit volume, procurement, security review, and anything the catalog cannot express. Route it to sales.

​
What a credit buys

A credit is the platform’s single unit of consumption. Every AI call, embedding, retrieval and ingestion job resolves into credits, so a customer watches one number instead of a provider invoice. The number that matters commercially is credits per conversation, and it has two useful values that must not be mixed up:
  • Typical (median): 49.5 credits. Half of conversations cost less than this. It is the right answer to “what does a conversation cost?”
  • Mean: 69.0 credits. It is higher because 29% of conversations run into the 150-credit ceiling and pull the average up. It is the right answer to “how many conversations does a month’s allowance buy?”, because an allowance is spent by a sum of conversations and the expected count of a sum is allowance ÷ mean.
Both come from the same measurement: 616 real conversations repriced under the current rate card. Quote the median when a prospect asks about cost, the mean when they ask about capacity — and never divide an allowance by the median, which over-promises by about 39%.

​
What a plan’s allowance buys

These are the figures the in-product upgrade page renders, from the same constant (CREDITS_PER_CONVERSATION_RATE_CARD_V2). Both are properties of the active rate card: activating a new card changes what a conversation costs, and this table, that constant and the numbers in the tier cards above must be re-derived together. Nothing will tell you they went stale. At the typical conversation: Say it plainly: a conversation costs about a quarter, and can never cost more than about seventy-five cents.

​
Against the alternatives

Intercom’s Fin charges $0.99 per resolved conversation. At a typical conversation, Brainstormer is 4× cheaper on Starter and 5× cheaper on Scale — and even a conversation that hits our internal ceiling is still cheaper than Fin’s flat rate. Zendesk’s AI resolutions are priced in the $1.50–2.00 range.
Those two competitor figures are the only ones we quote. Do not invent others. If a prospect asks about a vendor not listed here, say we have not benchmarked it rather than guessing — a made-up comparison is the fastest way to lose a deal in the second meeting.Also be precise about what Fin’s $0.99 counts: it is per resolved conversation, so a like-for-like comparison depends on resolution rate. Our figure is per conversation, resolved or not, which is the more conservative side of that comparison.

​
Why the bill is predictable

This is the retention story, and it is worth leading with. Under the previous rate card, conversation cost was effectively unbounded: the median conversation and the most expensive one differed by roughly 2,000×. A customer could have a good month and a catastrophic month with the same traffic, and nobody could forecast anything. The current rate card puts a hard ceiling on how much of a customer’s allowance a single conversation can consume. The spread collapsed: A customer on Growth can now say: 35,000 credits, at most 150 per conversation, so my worst case is 233 conversations and my realistic number is about 507. That sentence was not sayable before — the old card had no worst case — and it is the single strongest thing to put in front of a finance buyer. About 29% of conversations reach the ceiling, and that is the design working — the expensive ones get capped and everyone else pays less. It is also why the mean (69.0) sits well above the median (49.5), and why capacity must be quoted from the mean.

​
What burns credits beyond conversations

Sell this honestly. The surprise here is the main source of first-invoice complaints.
Adding a knowledge source means crawling it, chunking it, embedding every chunk, and extracting entities. It is real work with real provider cost, and it is charged.Ingestion is priced close to cost — it is onboarding, not the recurring value the subscription is for — and each ingestion job is capped. But the numbers are still material. On measured data, one large web source cost 6,000 credits to ingest. That is:
  • 75% of a Starter month
  • 17% of a Growth month
  • 5% of a Scale month
  • six times a Free month — a large source simply will not fit on Free
The largest source measured spanned thousands of documents and cost about 12,300 credits (35% of a Growth month).What to tell a customer: ingestion is a one-off cost per source, paid when you add it and again when you re-crawl. Budget for it in month one, separately from conversations. If they are bringing a large site on Starter, say so before they buy, not after.What NOT to promise: there is no per-source spending limit today. The caps bound each individual ingestion job, not the total for a source that spans a thousand pages. A customer who points us at an enormous site can consume an unbounded amount of their allowance. If a deal depends on a hard ingestion budget, escalate it — do not commit to one.
  • Conversation title generation — a fraction of a credit per conversation, charged outside the conversation cap.
  • Search reranking — charged per agent rather than per conversation, so it sits outside the per-conversation ceiling too.
  • Audio transcription, document summarisation, image and video understanding — charged when used, at close to cost.
  • WhatsApp Business messages — a flat 1 credit inbound, 2 credits per reply, on top of the AI cost.
  • Custom external tool calls — 1 credit per call.
Together these mean the real ceiling on a conversation is a little above the 0.73/0.73 / 0.64 / $0.58 figures above — call it “about seventy-five cents” and you will not be wrong.
Plans differ in credits, agents, knowledge bases and seats — not in capability. Every tier gets:
  • Access to 300+ models
  • Knowledge bases with document, web and social sources
  • Retrieval-augmented answers with citations
  • Multimodal input (images, audio, video, documents)
  • Prompt versioning and dynamic variables
  • API access
  • Organization roles and permissions
  • The full credit ledger, itemised
There is no “AI is a paid add-on” upsell and no feature paywall on the core product. That is a genuine differentiator against vendors who meter the model separately — lead with it.

​
Upgrade triggers

What actually makes someone outgrow a tier, in the order it tends to happen:
1

Free → Starter

The trigger is almost always ingestion, not chat. A serious knowledge base does not fit in 1,000 credits. The second trigger is the 1-agent / 1-knowledge-base limit — the moment they want a second audience, they need Starter.
2

Starter → Growth

Three signals, roughly in order of frequency:
  • Volume: consistently past ~116 conversations a month, or burning through the allowance before month end.
  • Sources that change: re-crawling a large site monthly costs a large fraction of a Starter allowance every time. Growth absorbs it.
  • Structure: they want more than 3 agents or 2 knowledge bases, or more than 5 people in the workspace. This is the cleanest upgrade conversation — it is a limit, not a forecast.
Growth is also where human handover becomes part of the story for teams that need an operator behind the agent.
3

Growth → Scale

Volume and predictability. Past ~507 conversations a month the per-credit saving alone (10% cheaper than Growth) starts to matter, and unlimited agents, knowledge bases and seats stop the customer having to ration structure to fit a plan.Sales-assisted today — see the note at the top of this page.
4

→ Enterprise

Procurement, security review, a committed volume that does not fit a published tier, or a contractual requirement the catalog cannot express.

​
Positioning cheat sheet

​
Frequently asked

Because customers do more than converse. Ingestion, transcription, summarisation and search all cost real money, and a per-conversation price either hides them (and we lose money on heavy users) or forces a second price list. One unit, one meter, one number to watch.The per-conversation cap gives the predictability of per-conversation pricing without pretending the other work is free.
Credit-consuming operations stop until the customer tops up or upgrades. See Overage and Invoices.
Plan credits are allocated per billing period. Add-on credits purchased separately do not expire. Check the customer’s billing page for their specific balance split before promising anything.
No, and they do not get a cut in allowance either. Both are frozen per account at signup. If you want a repricing to reach existing customers, ask an admin to apply it explicitly on the Plans page — it is a separate action with its own confirmation, because it changes what people already paying are entitled to.
Not below **0.00375percredit∗∗—thatistheflooratwhichtheplatformstopscoveringitsownprovidercostsontheheaviestworkloads,anditisenforcedinthedatabase.Scalesitsclosesttoit,at0.00375 per credit** — that is the floor at which the platform stops covering its own provider costs on the heaviest workloads, and it is enforced in the database. Scale sits closest to it, at 0.003838. Anything below the floor is an Enterprise conversation with a margin owner, not a discount a rep can apply.