BYOK for AI coding agents: what bring your own key actually costs

Every autonomous coding agent runs on somebody's model bill. The agent does its work by calling a frontier model thousands of times, and those tokens cost real money. So the question worth asking a vendor is not whether you pay for that usage. You always do. The question is how, and through whom. BYOK, bring your own key, is one answer: the AI coding agent runs on a key you hold directly with your model provider, and the tokens land on your provider invoice instead of the vendor's. Here's how the common pricing models compare, and when holding the key yourself is worth the trouble.

The three ways agent vendors charge

Strip away the branding and pricing collapses into three shapes.

Per-seat with bundled usage: you pay a flat monthly price per person, and some amount of model usage is included. The vendor buys tokens wholesale, sets the seat price to cover a typical user, and caps or throttles the heavy ones. It's predictable and easy to buy, and the cap is doing quiet work. A plan priced for the average user means light users subsidize heavy ones, and heavy users hit limits at the worst time.

Usage-based with a markup: you pay for what the agent consumes, but at the vendor's rates rather than the provider's. Sometimes it's framed as credits or "compute units", sometimes as a price per task. Underneath, the vendor pays the model provider one price and charges you another.

BYOK: you connect a key from your own model provider, or a subscription you already pay for. The vendor charges for the platform, usually per seat, and the model usage bills through your provider account at whatever rates you have there.

Why markups exist, and what they cost you

The markup model is not a scam, and it's worth being fair about why it exists. A vendor that resells usage carries real costs: it fronts the model bill, absorbs the occasional runaway task, deals with provider outages and rate limits, and gives your finance team one invoice instead of two. The margin on tokens funds all of that. For a team with no existing provider relationship and no interest in managing one, paying the margin can be a reasonable trade.

What it costs you is visibility and options. You can't see the spread between what the vendor pays and what you pay, and nothing stops it from changing. If your provider cuts prices, there's no guarantee the cut reaches you. Your negotiated rates, committed-use discounts, and cloud credits are useless here, because the usage runs on the vendor's account rather than yours. And your code goes to the model under the vendor's provider agreement, with data terms you may never have read.

What BYOK gets you with AI coding agents

The concrete case for BYOK comes down to four things.

You pay list price with zero markup. The token bill is your provider's published price, on your own invoice, itemized. There is no invisible margin between you and the model.

You can use the relationship you already have. If your company negotiated rates with a provider, or has committed spend or credits, agent usage counts against them. If you pay for a ChatGPT subscription, some agents (Sinatra among them) run on it directly, so a task costs you nothing extra until you hit the subscription's own limits.

You can switch models the day a better one ships. With bundled or marked-up pricing, you get the models your vendor has integrated and priced, on your vendor's schedule. With BYOK, a new model on your provider's API is a config change.

Your code flows under your own provider agreement. Whatever data-handling terms your team reviewed when it signed with the provider (retention, training, regional processing) are the terms that apply. You aren't inheriting a second set of terms you never negotiated.

The tradeoffs, honestly

BYOK moves work onto you. Someone has to create the key, add it to the agent platform, rotate it when policy says so, and watch the provider dashboard. Budgeting is yours too: there's no bundled cap to protect you from a heavy month, so you set the spending limits and alerts with your provider. Rate limits are also yours, which means an agent grinding through a backlog shares quota with everything else on the account.

And a bundled plan can genuinely be simpler for a small team. If you run a handful of tasks a week and nobody wants to own a provider account, one flat invoice may be worth more than the margin you'd save. BYOK pays off as usage grows, because a percentage markup grows with it and a flat platform fee doesn't.

How Sinatra prices it

Sinatra is BYOK on every tier. You bring an Anthropic key or an OpenRouter key, or connect a ChatGPT subscription you already pay for. Keys are stored encrypted and used only to run your tasks. The platform fee is $20 per member per month, with no markup on tokens on any tier; the free tier gives you 5 tasks a day on your own key or subscription, or 1 a day on Sinatra's key while you evaluate. The full breakdown is on the pricing page.

Model choice lives in a small config file in each repo, so pointing a repo at a new model is an edit and a commit. If you want to see where the model fits in the wider loop, the ticket-to-PR walkthrough covers it end to end.

Common questions

Is BYOK cheaper?

Usually at scale, but it depends on your rates. The structural difference is that with BYOK the platform fee and the token bill are separate and both visible, so you can see what a month of agent work cost and where. With markup pricing they're blended and the margin is invisible. If you have negotiated rates or committed spend with a provider, BYOK is the only model that lets you use them. A small team with light usage may find a bundled seat comparable in cost and simpler to run.

Can I use a ChatGPT subscription instead of an API key?

With Sinatra, yes. Connect the subscription you already pay for and tasks run against it, inside the subscription's own usage limits, with no separate API billing. On the free tier it counts the same as bringing a key: 5 tasks a day.

What happens to my key?

Sinatra stores it encrypted and uses it only to run your tasks. Because the key is yours, you can watch its usage in your provider's dashboard, and revoking it at the provider cuts off access immediately, on your schedule rather than anyone else's.

If you want to see what the loop feels like on your own key, you can start for free with 5 tasks a day, or read the docs for the setup details.