BYOK (bring your own key) means you run AI on your own provider API key, so each model call is billed at the provider's raw cost with no platform markup added on top. In ReplyLabs, you paste an OpenAI, Anthropic, Google, Mistral, or OpenRouter key into the add-on, and AI runs across your Google Sheets rows charge nothing to ReplyLabs for the model itself: the cost lands on your own provider account at the published rate. This is the same pattern coding tools and AI gateways adopted through 2025 and 2026, and it exists for one reason: at volume, a per-call markup on someone else's model is the most expensive line in your bill. This page explains what BYOK is, when it saves money, the exact pricing math against managed mode, which keys work, how keys are secured, and how to add one. For a one-paragraph definition see the BYOK glossary entry.
What is BYOK in plain terms
There are two ways a tool can pay for the AI it runs for you.
- Managed mode. The tool holds the provider relationship. It calls the model on its own key, pays the provider, and bills you a price that covers that cost plus a margin. You never see an API key. You top up a balance and the tool draws from it.
- BYOK mode. You hold the provider relationship. You create a key in your own OpenAI or Anthropic dashboard, paste it into the tool, and the tool calls the model as you. The provider bills your account directly. The tool charges nothing for the model call itself.
BYOK is the second shape. The platform still supplies everything around the model: the spreadsheet interface, the prompt templating, the row-by-row batching, the retries, and the result writing. What changes is who pays the model bill and at what rate. With BYOK there is no middle margin on inference, so you pay the provider's list price and not a cent more.
Why BYOK saves money at scale
A markup is invisible on a ten-row test and painful on a million-row month. Managed pricing is convenient precisely because it is bundled, but the bundle includes a margin on every single call. Once AI becomes recurring work rather than an occasional experiment, that margin compounds.
The break-even logic is simple. Managed mode adds a markup multiplier and a small per-row base fee. BYOK strips both away and leaves you with the provider's raw rate. The more rows you run, the more the stripped-away margin matters in absolute dollars. A team running a few hundred rows a week will barely notice the difference. A team running hundreds of thousands of rows a month is paying a recurring tax that BYOK removes entirely.
There is a second, quieter benefit: control. With your own key you see the spend in your provider dashboard, you can set provider-side budget alerts, and you can switch models the moment a cheaper or sharper one ships, without waiting for the platform to add it to a managed catalogue. You also avoid lock-in: the key is yours, the billing relationship is yours, and moving providers is a paste, not a migration.
BYOK vs managed pricing: the math
ReplyLabs prices managed AI at the provider's raw cost times 1.25, plus a base fee of $0.0025 per succeeded row. BYOK removes both: the model call runs on your key at the provider's raw rate, and there is no per-row base fee on the AI step.
Worked example. Suppose a given model and prompt cost the provider $0.0010 in raw tokens per row.
- Managed:
($0.0010 x 1.25) + $0.0025=$0.00375per succeeded row. - BYOK:
$0.0010per succeeded row, billed by your provider, with zero ReplyLabs charge on the AI step.
Across 100,000 succeeded rows that is roughly $375 managed versus $100 of provider spend under BYOK. The gap is the margin and base fee you stop paying. The cheaper the raw call, the larger the base fee looms in proportion, so high-volume runs on small fast models are where BYOK pulls furthest ahead. To plug in your own model rate and row count, use the AI cost calculator, which shows managed and raw figures side by side.
Two caveats keep the comparison honest. First, BYOK on AI is a Pro and Scale plan feature; the seat fee still applies, and that is where ReplyLabs earns its keep rather than on your inference. Second, only succeeded rows are charged in either mode, so failed or skipped rows cost nothing on the ReplyLabs side regardless of which mode you pick.
Which providers and keys work
ReplyLabs accepts AI keys from five providers, so you can match the key to the models you already use:
- OpenAI. Keys in the
sk-...format from your OpenAI dashboard. Covers the GPT family. - Anthropic. Keys in the
sk-ant-...format. Covers the Claude family. - Google. Gemini API keys from Google AI Studio.
- Mistral. API keys from the Mistral console.
- OpenRouter. A single OpenRouter key reaches hundreds of models across many providers through one unified endpoint, passing through each provider's published price with no inference markup of OpenRouter's own. This is the broadest option if you want one key to cover everything.
The practical upshot: if you want the widest model coverage from a single credential, an OpenRouter key is the simplest BYOK choice, because it fronts most major providers behind one key and one balance. If you already have a direct relationship and volume discount with one provider, use that provider's own key to keep billing in one place.
Email verification and scraping are separate steps with their own pricing and are not affected by your AI key. BYOK governs the AI step only.
How keys are secured
A provider API key is a credential that can spend money, so storage matters more than for almost any other field.
ReplyLabs encrypts every stored key at rest with AES-256-GCM, an authenticated encryption scheme that both conceals the key and detects any tampering with the stored value. The key is encrypted before it is written and decrypted only at the moment a call is dispatched on your behalf. It is never rendered back to the screen after you save it, never written to logs, and never exposed in run history or results. When you rotate or remove a key, the stored value is replaced or deleted.
The principle is least exposure: the key exists in plaintext only in transit to the provider at call time, and at rest it is ciphertext. That is the standard you should expect from any BYOK tool, and it is the one ReplyLabs holds itself to.
How to add a key in ReplyLabs
Adding a BYOK key takes under a minute:
- Create a key in your provider dashboard (OpenAI, Anthropic, Google AI Studio, Mistral, or OpenRouter). Scope it as your provider allows and, if the provider supports it, set a spend limit on that key.
- In Google Sheets, open the sidebar with Extensions, ReplyLabs, Open sidebar, then go to Settings, Connections.
- Paste the key into the field for the matching provider and save. The key is encrypted immediately and the field stops showing the raw value.
- Pick a model from that provider when you set up an AI run. The cost preview now shows the raw provider rate with no markup, confirming BYOK is active.
- Run as normal. The model call bills your provider account; ReplyLabs charges nothing for the AI step.
If you ever need to swap providers, paste a new key and select a model from that provider. There is no migration and no export.
When managed mode is simpler
BYOK is not always the right call. Managed mode wins when convenience outweighs the margin.
- Low or occasional volume. If you run a few hundred rows now and then, the saved margin is pennies and not worth holding a provider account.
- No existing provider account. Managed mode needs no key and no signup with OpenAI or anyone else. You top up a ReplyLabs balance and go.
- You want one bill. Managed keeps all spend, AI plus verify plus scrape, on a single ReplyLabs balance instead of split across provider invoices.
- You are still testing. During prompt design, managed mode lets you iterate without provisioning keys. Switch to BYOK once the workflow is stable and the volume justifies it.
A common path is to start managed, validate the workflow, and move the AI step to BYOK once monthly volume makes the margin visible. Because both modes run the same batching engine, switching is just a key paste in Settings. For the underlying mechanics of running prompts across a column either way, see AI in Google Sheets, and for help configuring a run see the AI prompts help guide.
Common questions
Does BYOK cost less than managed mode?
Yes, on the AI step. BYOK bills the model call at the provider's raw rate with no markup and no per-row base fee, while managed mode adds a 1.25x markup and a $0.0025 base fee per succeeded row. The savings grow with volume.
Is BYOK available on the free plan?
No. BYOK for AI is a Pro and Scale plan feature. New accounts get a $20 credit to try managed mode, and you can move to BYOK once you are on a paid plan.
Which AI providers can I bring a key for?
OpenAI, Anthropic, Google, Mistral, and OpenRouter. An OpenRouter key reaches hundreds of models across providers through one credential, which is the broadest single-key option.
Is my API key safe?
Keys are encrypted at rest with AES-256-GCM, decrypted only at call time, never shown again after saving, and never written to logs or run history. You can rotate or remove a key at any time.
Who pays the model bill under BYOK?
Your provider account does. The model call is made on your key, so the provider bills you directly at its published rate. ReplyLabs charges nothing for the AI step itself.
Can I switch between BYOK and managed mode?
Yes. Paste a key to use BYOK or remove it to fall back to managed. Both modes use the same batching engine, so switching is a settings change, not a migration. Start at the ReplyLabs home page to see the plans.