BYOK, or bring your own key, means you supply your own API credentials for the underlying AI models (the LLM that generates responses, the TTS engine that produces the voice) instead of routing all usage through the platform's own accounts. The platform still orchestrates the call, but the actual model usage bills to your account, at the provider's rate, not a marked-up reseller rate baked into your subscription.
Why does this matter for a voice AI platform specifically?
Every AI voice call runs on a stack of paid components: a speech-to-text pass, an LLM call to decide what to say, a text-to-speech pass to generate audio, and telephony minutes to carry the call. Platforms that don't offer BYOK typically bundle all of that into a per-minute or per-call price and mark up the model costs, sometimes heavily, because they're reselling access they've already paid for at volume.
With BYOK, you open your own account with the model provider (say, OpenAI for the LLM or a TTS vendor for the voice), generate an API key, and plug it into the platform's settings. The platform still handles orchestration, call flow, calendar integration, and telephony, but the model spend flows through your credentials and shows up on your own provider invoice, not folded into the platform's fee.
What does BYOK NOT do?
It doesn't make the platform free. You're still paying for the orchestration layer: the software that manages call flow, transcription pipeline, integrations, and infrastructure. BYOK removes the markup on model consumption, not the platform fee itself. It also doesn't eliminate technical setup, you still need to generate keys, manage rate limits, and handle what happens if a key expires or a provider account gets suspended for billing reasons. If you're not comfortable owning that operational surface, a platform that manages model access for you (at a built-in markup) may still be the simpler choice.
It's also not universal. Some platforms only support BYOK on higher-tier plans, or only for certain providers, or don't support it at all and price everything as an all-in per-minute rate. Vapi's own pricing page breaks out model and infrastructure costs separately, which is one version of transparent usage-based pricing that isn't quite the same thing as full BYOK. Worth reading a competitor's actual pricing docs rather than assuming BYOK support, since the term gets used loosely.
How much does BYOK actually save an agency?
It depends entirely on call volume and which models you route through. At low volume, the savings are marginal, the markup on a few hundred minutes a month is a rounding error next to your subscription cost. At meaningful scale, running hundreds or thousands of minutes a month across multiple client accounts, model markup compounds fast, because you're paying it on every single call, not once.
The honest way to figure out whether BYOK is worth the setup effort for your agency is to run your actual numbers rather than guess. We built a BYOK savings calculator that takes your monthly call volume and model choice and estimates what you'd save by bringing your own key versus paying a bundled rate. It's worth five minutes before you decide either way.
Does BYOK affect voice quality or reliability?
Not directly. Voice quality is a function of which model and TTS engine you choose, not whether you or the platform holds the key. If a platform supports multiple TTS providers and lets you pick per-agent, BYOK just changes who's billed for that choice, not what the choice sounds like. Reliability is a separate question: if your own provider account hits a rate limit or your card on file expires, that's now your outage to fix, not the platform's. Some platforms handle this with automatic failover to a backup provider even on BYOK setups, worth confirming before you commit a client's live phone number to a single-key setup.
Is BYOK the same thing as white-label?
No, and this is a common mix-up. White-label is about branding, the platform runs under your agency's name, your clients never see the underlying vendor. BYOK is about billing, whose account the model usage runs through. A platform can be white-label without BYOK (you brand it, but pay the platform's bundled rate) or BYOK without white-label (you control model costs, but the platform's name shows through). The two features solve different problems and an agency evaluating platforms should ask about both separately rather than assuming one implies the other. Our comparison of Vapi covers how orchestration platforms and end-to-end voice AI platforms differ on this exact point, since the two categories bundle these features very differently.
Who actually benefits from BYOK in practice?
Agencies running enough client volume that model markup shows up as a real line item, and agencies that want direct visibility into what each client's calls actually cost in raw compute, independent of what the platform charges them. If you're reselling voice AI to multiple clients under agency pricing, BYOK gives you a cleaner cost basis for figuring out your own margin on each account, since you're not backing into "platform cost minus estimated markup" to price a client correctly. It's less useful if you're running a handful of low-volume clients where the subscription fee dominates total cost anyway.
If you're weighing this alongside other platform decisions before committing to a reseller model, it's worth working through the broader checklist on /agencies first, since key ownership is one factor among several (carrier isolation, white-label depth, support model) that determine whether a platform actually works for a reseller business rather than just an end user.