How AI Aggregators Work: One Subscription Across Five Labs
TL;DR: An AI aggregator holds its own accounts at OpenAI, Anthropic, Google, xAI and DeepSeek, sends your prompt to whichever of their APIs you picked, and returns the answer in one interface. That account custody is the whole mechanism, and it is what makes all AI models in one subscription possible: the aggregator is the labs' customer, so you get one bill instead of five. It routes to models rather than hosting them.
Key Takeaways
- An aggregator holds accounts with each lab and calls their official APIs on your behalf, so one login replaces separate accounts at OpenAI, Anthropic, Google, xAI, and DeepSeek.
- The aggregator, not you, is the labs' customer. That single fact decides the pricing model, the metering unit, and which features can and cannot be offered.
- Five distinct billing units are in use across the field: per token, dollar-denominated credits, abstract compute points, message caps, and a one-time license with your own keys.
- Routing is not hosting. Aggregators do not run frontier models on their own hardware, and any service claiming to is almost certainly serving a smaller open model under a frontier name.
- Perspective AI routes to every model in its catalog on one $14.99/mo plan and meters text per token, images and video per generation, and agents per operation.
Quick Answers
How do AI aggregators work?
An AI aggregator maintains its own accounts and API keys with each model provider. When you send a prompt, the aggregator forwards it to the provider you selected, receives the response, renders it in its own interface, and deducts the cost from your balance. You interact with one app and one bill; the aggregator interacts with five or more providers.
Do AI aggregators use the real models?
Legitimate aggregators call the providers' official APIs, so the model answering is the same model the lab serves in its own app. The differences are the interface around it, how history is stored, and how usage is metered. Services advertising unlimited frontier access for a few dollars a month are the ones to check, because API costs make that arithmetic impossible.
Are AI aggregators just wrappers?
A wrapper puts a new interface on one provider. An aggregator holds accounts with many providers, routes between them, normalizes their different response formats, and carries thread context across a model switch. The routing and context layer is the actual engineering, and it is why some aggregators can switch models inside one conversation and others start a new thread.
An AI aggregator is an application that holds its own accounts with several model providers and calls their APIs on your behalf. You send one prompt from one interface; the aggregator forwards it to OpenAI, Anthropic, Google, xAI, or whoever you picked, receives the response, and deducts the cost from a single balance. Perspective AI works this way, and so does every other platform in the model access category. The reason the category exists is that this arrangement is what lets one subscription replace five lab accounts, and consolidating five accounts into one is the only thing an aggregator sells that a single lab cannot. The mechanics matter because they decide what an aggregator can charge, what it can meter, and what it can never offer.
How Do AI Aggregators Work?
An AI aggregator forwards your prompt to a lab's official API using its own account, returns the answer in one interface, and bills you once. It routes to models. It does not host them.
That is the whole mechanism, and everything below is a consequence of it. The consumer-facing story is "all the models in one app". The engineering story is account custody: somebody has to be the labs' paying customer, and in this arrangement it is the aggregator rather than you.
The Request Path, End to End
A single message crosses four boundaries between your keyboard and the answer on screen.
- Selection. You choose a model, or the platform chooses one for you. This is the only step where a router does anything interesting, and it is covered in how model routing decides which model answers.
- Translation. Your thread is rebuilt in the target provider's request format. Providers disagree about how system instructions, tool calls, images, and reasoning parameters are expressed, so the aggregator normalizes them. This is the step that makes a mid-thread model switch possible: the conversation is stored in the aggregator's own shape and re-serialized per provider on every turn.
- The call. The request goes to the provider over the aggregator's API key. The provider sees the aggregator as the customer. It does not see your subscription, your plan, or your account.
- Metering and return. The response streams back, the aggregator counts what it cost, deducts it from your balance, and renders the answer.
Step two is where the engineering lives, and it is worth dwelling on because it is the step people assume is trivial. Providers do not agree on much: system instructions are a separate field for one and a message role for another, image inputs arrive as different encodings, tool definitions use different schemas, and reasoning parameters exist on some models and not others. An aggregator therefore keeps the conversation in its own canonical form and rebuilds the provider-specific request every single turn. That is the reason a thread can survive a model change on some platforms and not others: where the thread is stored in the shape of whichever provider last answered, handing it to a different provider means starting again.
Two things follow immediately. First, latency is the provider's latency plus one extra network hop, which is small relative to the time a model spends generating. Second, whatever the provider does not expose over its API cannot appear in the aggregator, no matter how well built the aggregator is. That constraint is large enough to have its own page: what routing to a model cannot give you.
The One Subscription That Replaces Five Lab Accounts, and Who Holds the API Key
There are only two arrangements, and every platform in this market is one or the other.
The platform holds the keys. You pay the platform a subscription or buy credits from it. The platform pays the labs. You never see an API dashboard, never manage a key, and never get a surprise invoice from OpenAI. Perspective AI, Poe, and most consumer aggregators work this way. So does the cheap end of the field, where a bundle such as GLBGPT resells proxied access to 100+ models from about $5.80/mo, and our GLBGPT alternative guide sets out what the metered credit pool underneath that headline actually buys.
You hold the keys. The platform is software you point at your own provider accounts. You pay the labs directly at token rates and pay the platform once, or not at all. TypingMind is the clearest example: its documentation describes it as a chat interface where you "bring your own API keys (BYOK)", with a license purchased separately from any model spend.
Custody also settles a question people ask in the wrong order: whose rate limits apply. Under the first arrangement they are the aggregator's, pooled across its whole user base and managed by it, which is why an individual user rarely meets a provider-side limit but can meet a platform-side one. Under the second they are yours, tied to your own account tier at each lab, which is a very different failure mode and one you have to administer personally.
The first arrangement is why a subscription can undercut the plans it replaces. A lab's consumer plan is a flat fee sized for a heavy user and sold to everyone, while API access is priced per token. An aggregator buying at token rates across a pooled user base can charge one subscription and still cover its costs, because most users are not heavy users. The trade is that your usage now has a meter on it, which is the honest cost of the arrangement and the reason credits exist as a unit at all.
Five Meters for One Product Category
Nothing in this market has standardized on a billing unit. Checking the published plans of the main platforms on 18 August 2026 turns up five genuinely different meters, which is why cross-platform price comparison is so awkward:
| Meter | Who uses it | What you are counting |
|---|---|---|
| Per token | OpenRouter | Input and output tokens at each model's published rate, billed from a dollar balance |
| Dollar-denominated credits | Perspective AI, Venice AI | A credit unit with a stated relationship to money |
| Abstract compute points | Poe | Points per message, varying by model, with no published dollar value |
| Message caps | Single-lab plans, budget aggregators | Messages per window, with the window and the cap set by the vendor |
| One-time license plus your own keys | TypingMind | Nothing on the platform side; your provider invoices are the meter |
The meters are not interchangeable presentations of the same thing either. A per-token meter charges for length, so a habit of pasting long documents shows up immediately. A per-message cap charges for frequency, so the same habit is free until you run out of messages entirely. A points system charges for whichever of those the platform decided, at a rate you cannot convert. Someone whose work is a small number of very long analyses and someone whose work is a large number of short questions will get opposite answers about which plan is cheap, and both will be right.
Five meters for one category means the advertised monthly price is not a comparable number. Two plans at $18 and $19 can differ by an order of magnitude in what they let you actually do. The way through that is to price the output rather than the plan, which is what the price per answer index does, and to read the plan-by-plan cost of buying each lab directly before assuming a bundle saves money.
Aggregator, Router, and Wrapper Are Not the Same Thing
The three words get used interchangeably and describe different products.
- A wrapper puts a new interface on one provider. One upstream, one account, no routing decision to make.
- A router is infrastructure. It exposes many providers behind one API endpoint for developers to call from their own code, and bills per token. OpenRouter is the reference implementation: pay per token, no monthly fee, and its Auto Router will pick a model for you and price the response "at the same rate as the routed model".
- An aggregator is a consumer product built on that same idea: many providers, one login, one subscription, and a chat interface rather than an endpoint. Perspective AI covers its whole catalog on one plan from $14.99/mo and keeps thread context across a model switch, so the routing decision happens inside a conversation rather than between projects.
Priced against each other, for someone deciding which shape they need:
- OpenRouter: best for developers calling models from code, pay per token with no monthly fee.
- Perspective AI: best for replacing several consumer chat subscriptions, $14.99/mo.
- Claude Pro: best if one lab's app-layer features are the actual product you want, $20/mo.
- Venice AI Pro: best for a privacy-led single-vendor experience, $18/mo.
What Routing Means Inside a Perspective AI Thread
The abstract version of this is a diagram. The concrete version is one conversation where the model changes and nothing else does. You start a long document analysis with a large-context model, hand the same thread to a strong writing model for the draft, and finish with a fast cheap model for the formatting pass. Each turn is a separate API call to a separate company under Perspective AI's keys, metered per token as it happens, drawn from one balance you can watch move.
What the mechanism buys you is not magic. It is the removal of five signup flows, five invoices, and the copy-paste tax of moving context between apps by hand. Open a thread on Perspective AI and switch models inside it once; the architecture stops being abstract immediately.
FAQ
How do AI aggregators work?
An AI aggregator maintains its own accounts and API keys with each model provider. When you send a prompt, the aggregator forwards it to the provider you selected, receives the response, renders it in its own interface, and deducts the cost from your balance. You interact with one app and one bill; the aggregator interacts with five or more providers.
Do AI aggregators use the real models?
Legitimate aggregators call the providers' official APIs, so the model answering is the same model the lab serves in its own app. The differences are the interface around it, how history is stored, and how usage is metered. Services advertising unlimited frontier access for a few dollars a month are the ones to check, because API costs make that arithmetic impossible.
Are AI aggregators just wrappers?
A wrapper puts a new interface on one provider. An aggregator holds accounts with many providers, routes between them, normalizes their different response formats, and carries thread context across a model switch. The routing and context layer is the actual engineering, and it is why some aggregators can switch models inside one conversation and others start a new thread.
Why do aggregators cost less than paying each lab directly?
Because API pricing and consumer subscription pricing are different products. A lab's $20/mo plan is priced for a heavy user and sold to everyone; API access is priced per token. An aggregator pays token rates across a pooled user base and charges one subscription, so a user with normal volume pays less than the sum of the plans they replaced.
Why choose one AI when you can use them all?
Perspective AI gives you ChatGPT, Claude, Gemini, and more in one app.
Try Perspective AI →