How AI Credits Work: The Unit That Makes One Subscription Cover Every Model

TL;DR: AI credits are prepaid units a platform deducts per action instead of per month. Text usually meters by token, images per generation, agents per operation. Some platforms publish a dollar conversion for the unit, others use abstract points. The unit only matters because it is what makes one subscription across many model families possible: a flat per-lab plan cannot price a GPT answer and a DeepSeek answer differently, and a credit can.

You send a message, and a number moves. That is the whole of credit-based AI pricing, and the reason it is worth understanding is that the number moves by wildly different amounts depending on choices you are making without seeing them. It is also the mechanism that lets one subscription replace several single-lab plans: consolidating five vendors behind one bill only works if the bill can charge a cheap model less than an expensive one, and a credit is how that is done. Credits are now the default meter across multi-model platforms, including the credit system on Perspective AI, and they sit at the centre of the pricing question because they are the only unit that makes the cost of an individual AI action visible at all.

How Do AI Credits Work?

AI credits are prepaid units a platform deducts per action rather than per month. Text meters by tokens sent and received, images and video per generation, agent runs per operation. The balance falls as you use it.

The Three Things a Credit Actually Meters

Every credit system in this market meters three families of work, and it meters them in three different ways because the underlying providers do.

• Text and reasoning: per token. A token is roughly three quarters of a word. Both directions count, so what you send is charged as well as what comes back. This is the only one of the three where you control the bill through habit rather than through choice of feature.
• Images and video: per generation. A flat unit cost per output, usually varying with resolution, duration, or model. There is no length to economise on, only quantity.
• Agents and tools: per operation. Each step an agent takes is its own billable event: a tool call, a search, a command run. A single instruction can therefore expand into many operations, which is the one place where the cost of an action is genuinely hard to predict in advance.

The word "credit" is doing a lot of work across those three, and it is worth being clear that it is a wrapper rather than a measurement. Nothing about a credit is intrinsic. It is an accounting entry the platform creates so that three incompatible upstream billing schemes can be presented as one falling number. Whether that wrapper is useful depends entirely on whether it stays tied to something real, which is the question the rest of this page keeps returning to.

The reason platforms do not flatten all three into a single per-message fee is that the flattening is always wrong in both directions. A per-message price high enough to cover a video generation would savagely overcharge someone asking short questions, and one low enough to be fair to them would lose money on every agent run. So the meter follows the work.

Why the Same Prompt Costs Different Amounts

Send the identical question to two models and the deduction can differ by an order of magnitude. Three variables explain nearly all of it.

The model. Provider token rates are not close to each other. Efficient small models are built to be cheap per token; frontier reasoning models spend additional tokens thinking before they answer, and charge for those too. The gap between the cheapest and dearest model available on a given platform is normally larger than the gap between two competing platforms' subscription prices, which is why model choice moves your monthly spend more than plan choice does.

What you send. Input is charged. A pasted hundred-page report is the largest line on that message's bill before the model has written a word, and if the thread continues, the whole conversation is resent as context on every subsequent turn. Long threads therefore get more expensive per message as they go on, which surprises almost everyone the first time they watch it happen.

What comes back. Output is usually charged at a higher rate than input. Asking for a summary is cheap; asking for a full draft is not.

A fourth variable belongs here even though it is not always presented as one. Reasoning-capable models can be asked to think harder before answering, and the thinking is generated tokens like any other. Two identical prompts to the identical model at different effort settings can differ several times over in cost, which is why a bill can move without any visible change to what you did. Where a platform exposes that as a control, it is one of the largest levers you have; where it does not, it is a source of unexplained variance in your spend.

Once those three are visible on screen, a routing instinct forms without any effort: cheap fast models for the routine turns, expensive models for the turns that deserve them. That instinct is the practical payoff of a visible meter, and it is what choosing between automatic and manual model selection is really about.

Dollar Conversions and Abstract Points: A Six-Platform Check

The interesting variation between credit systems is not how they meter. It is whether the unit means anything in money. Reading the published pricing pages of six metered consumer AI plans on 18 August 2026:

Three of six publish a conversion between their unit and money. One meters in a unit that cannot be converted, and two do not meter on the platform side at all. That split matters more than it looks, because a unit with no dollar value cannot be compared with anything: 660,000 points a month is not more or less than 250 credits a month, and neither is comparable to a message cap. An abstract unit is not dishonest, but it does make the platform the only party who can price its own product, and publishing a per-action rate in that unit only moves the problem: Krater states a chat message at 1 to 5 credits and a video at 50 to a few hundred without stating what a credit is worth, which is the gap behind the Krater AI alternative case. A published rate is also not the same as a published allowance, which is the distinction behind what Venice AI credits actually buy at each tier.

It is worth separating two things that get confused here, because a dollar conversion is necessary and not sufficient. Knowing that a hundred of something equals a dollar tells you how to read a balance. It does not tell you what an action will cost before you take it, and those are different pieces of information serving different decisions. A conversion rate turns your balance into a budget. A per-action price turns each message into a choice. A system that publishes the first and not the second still leaves you finding out afterwards, which is a receipt rather than a decision, and it is the distinction the four questions further down are built to expose.

The second thing a published conversion enables is auditing the platform rather than just your own usage. If a credit has a stated dollar value and a model's rate is stated in credits, then the platform's markup over the provider's published token price is calculable by anyone who cares to do the arithmetic. That is uncomfortable for a platform whose margin varies by model, and comfortable for one pricing at cost plus a flat margin, which is a fair explanation of why the abstract-unit approach persists in this market.

A dollar-denominated unit collapses that problem. If a credit has a stated relationship to money, then the cost of an action is a price, your monthly spend is an addition, and comparison against the market price of each plan a credit balance replaces becomes arithmetic rather than guesswork.

Credits Versus Message Caps

The alternative to a meter is a ceiling, and most single-lab consumer plans use one. The difference runs deeper than the accounting.

Neither model is inherently better value. A cap can be excellent value if you are a heavy user of one model and never reach it. What a cap cannot do is tell you where you stand before you get there, and that is the practical complaint behind most of the frustration documented in the comparison of AI usage limits across plans.

What Drains a Balance Fastest

In descending order of impact for a typical month:

1. Video generation. Nothing else is close. A handful of clips can outweigh a month of chat.
2. Agent runs with tool use. Per-operation billing multiplies quietly. An agent that searches, reads, and retries is spending on every one of those steps.
3. Long documents in long threads. The context resend on every turn is the compounding one, and it is invisible unless the meter shows it.
4. Defaulting to the most expensive model. The single easiest thing to fix, and the one most people only fix once they can see it.
5. Image generation. Real but usually bounded, because you make images deliberately.

Two habits follow from that list and they are worth more than any plan change. The first is input discipline: trimming a pasted document to the pages that matter roughly halves the input cost of every turn in the thread that follows, because the whole conversation is resent each time. The second is starting fresh threads for unrelated work, since a long-running thread carries its entire history as billable input into a question that has nothing to do with it.

Four of those five are behaviour rather than plan choice, which is the argument for a meter you can watch. You cannot economise on a number you never see.

Four Questions That Expose Any Credit System

Before buying into any credit-metered plan, ask these in order. They take about five minutes on the vendor's own pages and they separate the systems that are simply metered from the systems that are opaque.

1. What is one unit worth in dollars? If the page cannot answer this, you are being asked to budget in a currency with no exchange rate.
2. Is the cost of an action shown before I take it, or only after? After is a receipt. Before is a decision.
3. How is the per-model rate set, and what happens when a provider cuts prices? A cost-plus rate passes cuts through. A fixed markup per model keeps them.
4. What happens at zero? Top-up, hard stop, or silent downgrade to a cheaper model are three very different products.

Where a Balance Comes From

Three sources are standard across the field, and the differences between platforms are mostly in the fine print rather than the structure. A recurring plan allowance arrives on the billing date. Top-ups are bought on demand and usually behave differently from allowance credits on expiry, which is the detail worth checking. Promotional and referral credits are the third, and they are the ones most likely to carry restrictions on which models or features they apply to.

The healthy pattern is one pool with one price sheet: allowance, top-up, and bonus credits all spending identically. The pattern to watch for is segregated balances with different rules, because that is where a headline allowance can turn out not to cover the thing you wanted to do.

Does One Subscription's Credit Pool Replace Five Separate Bills?

Concretely, on our own system: Starter is $14.99/mo, Pro is $49.99/mo, and Enterprise is custom priced. Each plan carries a monthly credit allowance, and the allowance is the only thing that changes between them, because the same balance covers every model in the catalog. Text meters per token, images and video per generation, and agents per operation, with each model priced at its real provider cost plus a thin platform margin rather than a per-model markup. Every action shows its cost before your balance moves. That is what makes the consolidation arithmetic checkable: one balance drawn against five labs, instead of five plans you have to reason about separately. The full detail, including how referral and top-up credits behave, is on our own credits page.

The generic point survives without us, though. A credit system is only as useful as its dollar value is knowable, and that single test sorts this market faster than any feature comparison. Watch a balance move on Perspective AI and you will know within one session what the meter is actually charging you for.

FAQ