How credits actually work
Every AI reply costs credits, and the cost depends on which model produced it. That is the whole model — there is no per-seat fee, no charge for starting a conversation, and no cost for reading one back later.
| Tier | Credits per reply | A 3-model round | An 8-reply debate |
|---|
| Economy | 5 | 15 | 40 |
| Standard | 8 | 24 | 64 |
| Premium | 12 | 36 | 96 |
File attachments cost 2 credits each on top, and every participating model sees the attachment.
Which plan fits
Starter at $19 gives you 800 credits a month and up to 2 models per conversation. On economy models that is roughly twenty full debates; on premium models, closer to eight. Two models is genuinely enough for most questions — a clean argument between two competent models is easier to follow than a three-way one.
Pro at $49 gives you 3,000 credits and a third model per conversation. The third seat earns its cost specifically when you suspect you are asking the wrong question, because reframing needs an existing disagreement to push against. If you can already articulate the tradeoff you are weighing, Starter is the better buy.
Credit packs never expire and stack on top of your monthly allowance. Subscription credits reset each renewal; purchased credits carry over indefinitely, and your monthly allowance is spent first.
Spending less without getting less
The instinct on an important question is to reach for the most expensive models available. It is usually the wrong call.
Across every family, the gap between a flagship and a current-generation mid-tier model is much smaller than the gap in price. On most questions the flagship does not produce a better argument — it produces a longer one with more qualifications. Flagships earn their price on long chains of dependent inference, where an early mistake invalidates everything after it.
A configuration that works well and costs 21 credits a round rather than 36: two standard-tier models from different providers, plus one economy model from a third as the outsider. The economy model will not out-argue the others, but it changes what they have to answer — and that is most of what a third participant is for.
Worth knowing: tier follows current provider cost, not release date. Model pricing falls over time, so a newer model is sometimes cheaper than the one it replaces. Buying “the newest” and buying “the most expensive” are different decisions, and conflating them wastes money.
Comparing this to separate subscriptions
Individual subscriptions to the major assistants run around $20 a month each. Three of them is roughly $60, and you pay full price for each whether you use it daily or twice. Usage-based pricing changes the shape of that: a model you consult occasionally costs you occasional money rather than a standing fee.
The honest counter-case: heavy daily use of one specific model is usually cheaper on that model's own plan, and comes with product features — custom assistants, project workspaces, desktop integrations — that a usage-based platform does not replace. If you live in one assistant all day, pay that vendor directly.
Where this wins is spread and burstiness. Several models, or a lot of usage one week and none the next. And it wins outright on the thing no combination of separate subscriptions can buy at any price: having the models read and rebut each other in one conversation.
Common billing questions
Do unused monthly credits roll over? No — subscription credits reset on renewal. Credit packs are different: those never expire and carry over indefinitely. Your monthly allowance is always spent first, so packs act as a reserve rather than being burned on routine use.
What happens if I run out mid-debate? The conversation stops and stays saved. Top up with a pack or wait for renewal, then reopen it from the dashboard and continue where it left off. Nothing is lost.
Can I change plans? Yes, from the billing page, and changes are prorated by Stripe. Downgrading reduces how many models you can put in one conversation, but existing saved conversations remain readable and exportable regardless.
Is there a free trial? New accounts get free trial credits for the economy models, no card required. That is enough for a few complete debates — enough to find out whether reading two models disagree is useful to how you think, which is the only question worth answering before paying.
Do you charge for failed requests? No. If a model errors or times out, no credits are deducted.
Worked examples
Abstract credit arithmetic is hard to reason about, so here are three realistic sessions with real totals.
A quick two-model sanity check. Two economy models, four replies each way, one interruption from you. Eight model replies at 5 credits each: 40 credits. On Starter that is one-twentieth of your monthly allowance, and it takes about three minutes.
A serious three-model decision. Two standard models plus one economy outsider, three full rounds, two interruptions. Six standard replies at 8 and three economy at 5: 63 credits. Under Starter's allowance you could run twelve of these a month.
A hard question, no compromises. Three premium models, four full rounds, a document attached. Twelve premium replies at 12 plus one attachment at 2: 146 credits. On Pro that is roughly twenty such sessions a month; on Starter, five.
The pattern worth noticing: mixing tiers is not a compromise. The 63-credit session above frequently produces a better discussion than the 146-credit one, because the economy model contributes provider diversity rather than raw capability — and diversity is what generates disagreement.
Why usage-based rather than unlimited
Unlimited plans sound better and behave worse. Every provider offering one either caps you quietly once you become expensive, or routes you to a cheaper model without saying so. Both mean the plan you thought you bought is not the plan you have, and you find out at the moment you most need capacity.
Charging per reply keeps the arithmetic visible. You can see what a model costs before you pick it, see the balance move as a discussion runs, and decide whether a given question is worth a premium participant. That is a slightly less comfortable purchase and a considerably more honest one.
It also aligns the incentives properly. We have no reason to quietly downgrade the model you selected, because we are not absorbing its cost — you chose it and you are paying for it. Every reply comes from the model named on the message.
One consequence worth stating: because you pick the model and pay for that model, a reply attributed to Claude Opus 5 came from Claude Opus 5. Some platforms route requests to whatever is cheapest at that moment and label the output generically. If you are comparing how models reason, silent substitution destroys the entire exercise, so we do not do it.
What is not charged for
Reading a saved conversation, exporting one as Markdown, text or PDF, switching between Free Talk and Structured mid-conversation, pausing, interrupting, or changing the debate mode. You pay for model replies. Everything around them is included.
Credits are also not consumed by a failed request. If a model errors or times out, nothing is deducted.