How to Use Your AI Subscriptions as an API

Count what you pay for AI each month. For a lot of people the answer is somewhere between twenty and sixty dollars, spread across two or three services, because the models are genuinely different and you have learned which one to ask for what.

Now try to do something programmatic with any of them — summarise a folder of documents, run a prompt over every row of a spreadsheet, put a feature into a side project. Every provider gives you the same answer: open a second account, the one that bills per token, on top of the subscription you are already paying for. You are not buying access at that point. You are buying it twice.

That gap is the subject of this post: why it exists, what it actually costs you, and the specific trade you can make to close it.

A subscription and an API are not the same product

It is worth being precise about why the two bills exist, because the reason is not greed and it changes what a workaround has to look like.

A subscription buys you a web application. Rate limits are loose, fair-use policies are vague, and the provider assumes a human is on the other end — reading, thinking, typing again. A model API buys you metered inference. It is billed per token because that is genuinely what it costs to serve, and it is sold with a latency and reliability profile that a web app never promises.

So they are different products, priced differently, for good reasons. The problem is not the pricing model. The problem is that most people's programmatic needs are nothing like production inference:

• A script that runs overnight and does not care about latency.
• A batch of two hundred prompts, once, to see whether an idea works.
• A comparison run: the same question to four models, so you can read the disagreement.
• A hobby project with no users, which needs a card on file anyway.

None of that needs millisecond responses or five-nines uptime. All of it gets billed as though it does — and worse, all of it requires you to open a metered account and then keep half an eye on it, because a loop with a bug in it is now a loop that spends money.

The other half: nobody uses one model any more

The second thing that has quietly changed is that model loyalty is gone. People keep several subscriptions on purpose — one for long-context reading, one for code, one for search-grounded answers — and switch between them by habit.

Which means the daily reality of using AI well now involves a lot of manual work that nobody designed: open three tabs, paste the same prompt into each, wait, then read three answers one after another and try to hold them in your head at the same time.

That is not comparison. Comparison means seeing where the models disagree, and you cannot see disagreement in answers that arrived four minutes apart in three different windows. By the time you read the third one you have already been anchored by the first. If you have ever tried to decide whether one model is genuinely better than another for your particular workload, you know this failure mode intimately — and you know how often it ends with a shrug and sticking with what you had.

The workaround: drive the accounts you already have

There is a third option between "pay twice" and "give up", and it is unglamorous: instead of calling a provider's API, drive the provider's web app — the one your subscription already entitles you to — in a real browser, on your own machine, signed into your own account.

This is the idea behind Revoye, the AI provider router we have just released. The mechanic is simple enough to state in one sentence: each agent is a browser tab logged into one of your AI accounts, and a router in front of that fleet decides which tab gets which prompt.

From there you get two things that were previously separate purchases:

1. A multi-agent chat. One prompt goes to every enabled account at the same time, and the answers come back in one thread or in side-by-side columns. Because they run in parallel, six answers take roughly as long as one.
2. An API. The same fleet sits behind a single OpenAI-compatible endpoint, so your code can post a prompt and get a completion back — over the subscriptions you already hold, with no provider API key anywhere in the system.

Revoye currently works with ChatGPT, Claude, Gemini, DeepSeek, Qwen, and Perplexity.

Revoye is not just a browser extension

This is the part that matters most for understanding what you are actually installing, and it is where the mental model usually goes wrong. Revoye is one product with three interfaces, not three products:

Revoye Web — the router

The hosted half. It holds your account, issues and revokes API keys, pairs your devices, keeps the job queue, applies your routing rules, and shows a status dashboard covering every device, provider, and agent you have connected. You reach it in an ordinary browser tab at revoye.degird.com.

The browser extension — the part you touch daily

It runs the agents, drives the provider pages, and hosts the chat window. It works in Chrome, Edge, Brave, and other Chromium browsers from version 111 onwards — including on Linux.

It is also the only piece you need for the chat half. No account, no desktop install, nothing else. That is deliberate: most people meet Revoye by installing the extension, pointing three agents at accounts they already have, and asking all of them the same question.

Revoye Desk — your machine's identity

A desktop app for Windows and macOS, needed only for the API half. Its job is cryptographic identity: it generates a keypair for the machine, keeps the private half in the operating system's own keychain — Credential Manager on Windows, Keychain on macOS — and holds an authenticated connection back to the router so API work can be handed down to the extension. You install it once and then largely forget it.

The three share the same fleet of agents. An agent busy answering a chat prompt is simply skipped by the API, rather than being asked to answer two things at once.

Routing: deciding which account gets which prompt

Once you have more than one agent, "which one answers?" becomes a real question, and Revoye gives you four strategies rather than one opinion:

Strategy	What it does
Round robin	Rotates evenly through agents and providers
Least recently used	Picks whichever agent has been idle longest
Priority	Prefers a named provider, falls back when it is unavailable
Time window	Keeps a provider to the hours you configure

On top of that you can set hourly rate limits per agent, per provider, and globally — which is what stops a runaway script from spending a whole day's headroom on one account in ten minutes — and you can request a specific provider per call or let the router choose.

What it costs you: the honest limitations

Every workaround is a trade, and this one has a specific shape. It is better to know it now than to find out on your third integration.

• It is slow. Roughly 20 to 90 seconds per prompt, because a browser is typing into a page and a model is generating an answer into it. A model API answers in a fraction of that.
• There is no streaming. The agent reads a finished answer off the page, so there is nothing partial to forward. If your UI depends on token-by-token output, this is the wrong tool.
• The API half needs your computer online. Prompts queue durably when it is not — nothing is lost, and the queue drains when the device reconnects — but the answer waits.
• Providers redesign their websites. Agent automation can break when they do, until the agent is updated. Revoye does not spoof user agents or manipulate browser fingerprints to get around this; it drives the pages as the browser presents them.
• One owner per router. No shared workspaces, no roles, no seats in this version.

Where your logins actually live

The obvious question about a product that uses your ChatGPT account is where your ChatGPT login ends up.

The answer is that it does not move. Your provider sessions stay in your own browser, exactly where they already were. Revoye stores no provider password, cookie, or session token, and the extension has no access to them either — it talks to Revoye Desk over a local bridge. The keys Revoye does issue are scoped, revocable, shown to you once, and stored on the server only as hashes.

What Revoye does store is worth stating too, because "we store nothing" is almost never true: your email and a password hash, API key hashes, device public keys, your routing configuration, and API job history — prompts, responses, timings, and which agent answered — which is what makes asynchronous retrieval work at all. That history is removed on a retention schedule and you can delete jobs yourself. Conversations in the chat window never leave your browser. The full detail is in our privacy policy.

If you already have OpenAI SDK code

The endpoint is OpenAI-compatible on purpose, because the cost of adopting a new AI backend is almost never the backend — it is the integration you have to rewrite around it.

In practice, an existing integration needs a changed base URL and a model name of revoye/chatgpt or revoye/auto, and nothing else. Requests can wait inline for an answer or return immediately and call a webhook when the result is ready, and idempotency keys mean a retried request cannot quietly become two jobs. The Revoye developer documentation covers the endpoint, errors, rate limits, webhooks, and the exact edges of OpenAI compatibility.

Getting started

There are two ways in, and the first one is a single step.

1. For the chat: install the browser extension, create a couple of agents pointing at accounts you are already signed into, and send a prompt. That is the whole setup — no Revoye account required.
2. For the API: create a free account, install Revoye Desk on the Windows or macOS machine that will do the work, pair it with the one-time code from your dashboard, generate a scoped API key, and post your first completion.

The honest recommendation is to start with the chat, because it costs an install and answers the question most people actually have: is the model I have been using the right one for this particular job? You can settle that this afternoon, against your real prompts, with the subscriptions already on your card.

The full breakdown — every component, every routing strategy, the complete feature list, and the limitations in one place — is on the Revoye product page. If you are curious how it sits alongside the rest of what we build, the Degird product portfolio is a reasonable next stop, and our take on why Claude is leading the AI industry is a decent companion piece for anyone deciding which subscriptions to keep in the first place.