BYOK explained: bring your own Anthropic, OpenAI, or Bedrock key

What BYOK means

BYOK stands for "bring your own key." Instead of buying AI usage through a SaaS vendor that resells model access, you create an API key directly with the model provider — Anthropic, OpenAI, or AWS Bedrock — and connect that key to the platform. From then on, every call the platform's AI agents make runs on your key. The model vendor bills you directly, at their published rates, on your account.

The alternative most AI products ship with is the opposite arrangement: the vendor owns the model account, runs all customer traffic through it, and charges you for "AI credits," "messages," or a usage tier. That's convenient on day one, but it inserts a reseller between you and the thing actually doing the work. You pay their margin, your data flows through their account, and your usage history lives with them.

BYOK removes that middle layer for the one input that matters most in an AI product — the model. The platform still provides everything around the model: the industry template, the agent team, the connectors to your CRM, inbox, and codebase, the human-in-the-loop approval flow. But the raw inference is metered against your own vendor relationship.

Why it matters: cost control

The headline reason founders care about BYOK is price transparency. Model inference is cheap and getting cheaper, and per-token rates are published. As of now, a million input tokens costs about $1 on a fast, small model (Claude Haiku 4.5), around $3 on a mid-tier model (Claude Sonnet 4.6), and $5 on a top-tier model (Claude Opus 4.8). Output tokens run a few times those numbers. A million tokens is a lot — roughly 750,000 words of input — so a solo founder's real monthly model spend usually lands under $50.

When a reseller sits in the middle, you don't see those rates. You see a packaged price designed to cover the vendor's cost plus margin plus the risk that you'll use more than average. With BYOK, the margin is gone. You pay $5 for what costs $5, and you can watch it accrue in your own Anthropic or OpenAI dashboard in real time.

It also means you control the cost levers directly. Choosing a cheaper model for routine work, leaning on prompt caching for repeated context, and capping usage are all decisions you make against your own account. As you scale, you negotiate volume terms with the vendor, not with a reseller who has every incentive not to pass the savings on.

Why it matters: data control

With BYOK, your prompts, documents, and the context the agents pull from your stack travel from your environment to the model vendor under your own account and your own data-retention configuration. You set whether the vendor retains data, for how long, and under which enterprise terms. Nothing is pooled through a shared third-party account where your data sits next to other customers' traffic.

This is the difference between "trust us with your data" and "your data, your vendor agreement." For a founder handling customer records, financials, medical content, or anything regulated, that distinction is the difference between a quick security review and a hard no. The platform sees what it needs to orchestrate the agents; the sensitive inference happens against credentials you own and can audit.

It also clarifies the trust boundary. A platform built around BYOK should store your key encrypted, never display it back to you, and scope it strictly to your own tenant. The model vendor's logs and dashboards become your system of record for what was actually sent — not a reseller's summary you have to take on faith.

Why it matters: no lock-in or markup

Resold AI creates two kinds of lock-in. The obvious one is the markup — you're paying above cost, and switching away means re-pricing your whole AI budget. The subtler one is that your vendor relationship, usage history, and any prepaid balance live inside someone else's account. Leaving is a negotiation.

BYOK inverts that. The model account is yours from the start. If you decide to leave the platform, you revoke one key in your vendor console and the platform's access ends. Your billing history, your data-retention settings, and your relationship with Anthropic or OpenAI all stay exactly where they were. There's no proprietary credit balance to forfeit and no reseller margin you were trapped paying.

It also future-proofs your model choice. Because you hold keys with the vendors directly, you can bring whichever provider fits — Anthropic, OpenAI, or AWS Bedrock — and change your mind later without the platform standing between you and the model market.

How it works in practice

Setup is a one-time step. You generate an API key in your model vendor's console — the Anthropic console, the OpenAI dashboard, or your AWS account for Bedrock — and paste it into the platform once. The platform stores it encrypted and uses it as the credential for every model call its agents make on your behalf. Standing the whole thing up, key included, takes about five minutes.

From then on it's invisible. When the CEO agent plans work and the office agents execute it — reading your CRM, drafting an email, proposing a code change — each underlying model call authenticates with your key. Because nothing fires without your approval, you stay in the loop on the actions while the inference quietly runs against your account. Usage shows up in your vendor dashboard the same way a call from your own code would.

Day-to-day management lives with the vendor, which is the point. Rotate the key, set spending limits, inspect usage, or tighten data-retention terms in the console you already control. The platform reflects whatever you set there. If you ever need to cut access cold, revoking the key in your vendor console does it instantly — no support ticket, no waiting on the platform.