Is BYOM completely free?

No — LLM tokens on your own endpoint are zero-rated in platform credits, but a platform-infrastructure fee still applies (15% of the default model's blended rate, on your actual token volume) to cover real compute, database and streaming costs.

What is the BYOM platform fee based on?

15% of the default model's blended per-token rate, applied to the same token volume your BYOM usage generates — not 15% of your own provider's bill.

When does BYOM actually save money?

When you already pay for spare model capacity elsewhere (an underused enterprise deployment) and your usage volume is high enough that the ~85%-discounted platform fee beats what metered credits on a comparable model would cost.

Can I use BYOM for only some agents?

See [the BYOM setup guide](/docs/byom) for exactly how per-agent switching and reverting works — this post focuses on the cost math, not the configuration steps.

Is metered credits ever cheaper than BYOM?

Often, if you're already on a very cheap default model — GPT-5.6 Luna's metered rate is low enough that BYOM's platform fee plus an external endpoint frequently doesn't beat just staying on metered credits, unless you have spare model capacity elsewhere already.

Product

Is Bring Your Own Model (BYOM) worth it? A real cost breakeven

TL;DR

Bring Your Own Model (BYOM) routes every agent's LLM calls straight to your own OpenAI-compatible endpoint instead of the built-in AI Gateway — but it isn't free. Agent Planners still bears real serverless compute, database and streaming costs for every run, so BYOM charges a platform-infrastructure fee equal to 15% of the default model's blended rate, applied to your actual token volume. That's the honest number this post runs the math on: when BYOM's 15% fee plus your own provider bill beats metered credits, and when it doesn't.

What BYOM actually changes

	Default (AI Gateway)	BYOM
Who you pay for LLM tokens	Agent Planners (metered in credits)	Your own model provider, directly
LLM tokens in Agent Planners credits	Metered at the model's real gateway rate	$0 — zero-rated
Platform-infrastructure fee	Included in the credit rate	15% of the default model's blended rate, same token volume
Tool, browser, sandbox, media calls	Metered normally	Metered normally — unchanged by BYOM

The platform fee exists because Agent Planners still runs the serverless compute, database writes, SSE streaming and workflow orchestration for every BYOM task — those costs are real regardless of which model answers the LLM calls.

The actual breakeven math

The platform fee is 15% of the default model's blended rate, not 15% of your own provider's bill. If the default model (GPT-5.6 Luna, priced at $0.20/$1.20 per 1M tokens) is the reference, the platform fee works out to roughly 15% of what metering Luna directly would have cost — meaning BYOM's fixed platform cost is already about an 85% discount versus running the same volume through the default metered model.

So the real question isn't "is BYOM free" (it isn't) — it's "is my own provider's per-token cost, plus that ~85%-discounted platform fee, cheaper than metered credits on whichever model I'd otherwise run." For a team already paying for a large enterprise model deployment they're not fully utilizing elsewhere, the marginal cost of routing Agent Planners' calls through it is often close to zero — making BYOM's total cost close to just the platform fee.

When BYOM tends to win

You already pay for a model deployment (an enterprise OpenAI-compatible endpoint, a self-hosted model) with spare capacity — the marginal cost of routing more calls through it is small.
Your usage volume is high enough that the 15% platform fee is clearly smaller than what metered credits on a comparable model would have cost.
You have a compliance or data-residency reason to keep LLM calls on infrastructure you control, independent of raw cost.

When metered credits tend to win

You're on a cheap default model already (GPT-5.6 Luna's metered rate is itself extremely low) — the 15% BYOM platform fee plus a comparably-priced external endpoint rarely beats just staying metered.
Your usage is low or spiky — you're not paying for spare model capacity elsewhere, so BYOM adds a fee without removing a cost you were already carrying.
You want per-agent model choice without operational overhead — BYOM applies to your whole organization's routing to one endpoint, while metered credits let each agent run a genuinely different model.

How to actually set it up

The setup itself — pasting your base URL, model id and API key on the Agents page, and how the platform test-calls it live before saving — is covered in Bring your own model (BYOM). This post is the cost decision; that page is the how-to.

Frequently asked questions

Is BYOM completely free?: No — LLM tokens on your own endpoint are zero-rated in platform credits, but a platform-infrastructure fee still applies (15% of the default model's blended rate, on your actual token volume) to cover real compute, database and streaming costs.
What is the BYOM platform fee based on?: 15% of the default model's blended per-token rate, applied to the same token volume your BYOM usage generates — not 15% of your own provider's bill.
When does BYOM actually save money?: When you already pay for spare model capacity elsewhere (an underused enterprise deployment) and your usage volume is high enough that the ~85%-discounted platform fee beats what metered credits on a comparable model would cost.
Can I use BYOM for only some agents?: See the BYOM setup guide for exactly how per-agent switching and reverting works — this post focuses on the cost math, not the configuration steps.
Is metered credits ever cheaper than BYOM?: Often, if you're already on a very cheap default model — GPT-5.6 Luna's metered rate is low enough that BYOM's platform fee plus an external endpoint frequently doesn't beat just staying on metered credits, unless you have spare model capacity elsewhere already.