Provider billing and gateway billing are separate
With provider-direct BYOK, OpenAI, Anthropic, Google or another upstream provider charges your provider account for inference. The gateway may separately charge a subscription, request fee, percentage fee, log fee or infrastructure fee.
The clean comparison is total platform cost plus provider cost—not whether a landing page says “zero markup.” A zero-markup gateway can still have a meaningful subscription or usage charge, and a percentage-based gateway can be cheaper for a very small workload.
Key custody is the real security question
A production gateway should make it clear where upstream keys live and what your application sends. One pattern is an encrypted server-side provider vault: applications receive a scoped gateway credential while the upstream key stays inside the control plane.
Another pattern forwards or references provider keys per request. That can be simpler, but it increases the number of systems and request paths that handle root provider credentials.
- Encryption at rest for stored provider keys.
- No raw-key return to the browser after storage.
- Scoped/revocable gateway credentials for applications.
- Per-key rate and spend limits.
- Clear deletion and rotation behavior.
BYOK and portability
BYOK reduces billing lock-in because your relationship with the model provider remains direct. It does not automatically eliminate gateway lock-in: routing rules, observability formats, virtual model aliases and agent workflows can still become platform-specific.
Prefer an OpenAI-compatible request surface and keep important routing configuration exportable. That makes a future migration less painful even if you never use it.