BYOK vs managed AI pricing: when each is cheaper

BYOK (bring your own key) is cheaper than managed AI whenever you run enough rows for the saved markup to outweigh the convenience of not holding a provider account. In ReplyLabs, managed AI bills the provider's raw cost times 1.25 plus a $0.0025 base fee per succeeded row, while BYOK bills the provider's raw rate directly to your own account with no ReplyLabs charge on the AI step. The two modes run the same batching engine, so the only real question is which one costs less for your volume, and the answer turns on the per-row maths below. This article shows that maths, the break-even point, and the cases where managed mode still wins despite costing slightly more per row.

What is the difference between BYOK and managed AI pricing?

The difference is who holds the provider relationship and what margin sits on each model call. In managed mode, ReplyLabs calls the model on its own provider key, pays the provider, and bills you a price that covers that cost plus a margin. In BYOK mode, you paste your own OpenAI, Anthropic, Google, Mistral, or OpenRouter key, ReplyLabs calls the model as you, and your provider invoices you directly at list price. Managed mode adds a markup and a base fee; BYOK strips both away. For a fuller definition see the BYOK glossary entry.

Everything else is identical between the two modes. The spreadsheet sidebar, the prompt templating, the row-by-row batching, the retries, and the result writing all behave the same. The only thing that changes is the price of the inference and who gets the bill for it.

The pricing maths, side by side

ReplyLabs managed AI uses two numbers: a markup of 1.25x on the provider's raw token cost, and a base fee of $0.0025 per succeeded row. BYOK uses one number: the provider's raw token cost, billed to your account, with zero ReplyLabs charge on the AI step.

So for a row whose raw model cost is R:

• Managed cost per row = (R x 1.25) + $0.0025
• BYOK cost per row = R (billed by your provider)

Worked example one. A small fast model on a short prompt costs the provider $0.0010 per row in tokens.

• Managed: ($0.0010 x 1.25) + $0.0025 = $0.00375 per row.
• BYOK: $0.0010 per row.

Across 100,000 succeeded rows that is $375 managed versus $100 of provider spend under BYOK. The $275 gap is the markup and base fee you stop paying.

Worked example two. A larger model on a longer prompt costs the provider $0.0080 per row.

• Managed: ($0.0080 x 1.25) + $0.0025 = $0.0125 per row.
• BYOK: $0.0080 per row.

Across 100,000 rows that is $1,250 managed versus $800 BYOK, a $450 gap. The absolute saving is larger on the expensive model, but the proportional saving is larger on the cheap one, because the fixed $0.0025 base fee is a bigger slice of a small raw cost.

To estimate the raw cost R for your own model and prompt, multiply input tokens by the provider's input rate and output tokens by the output rate, then divide by one million. You can plug both figures into the AI cost calculator, which shows managed and raw side by side.

When is BYOK cheaper than managed?

BYOK is cheaper on the AI step at every volume, because it removes a markup and a fee that are always positive. The honest question is not whether BYOK is cheaper per row (it always is) but whether the total saving justifies holding a provider account.

The saving on the AI step for N succeeded rows is:

saving = N x ((R x 0.25) + $0.0025)

At R = $0.0010 that is N x $0.0028. So:

• 1,000 rows saves about $2.80. Not worth provisioning a key.
• 10,000 rows saves about $28. Borderline.
• 100,000 rows saves about $280. Clearly worth it.
• 1,000,000 rows saves about $2,800 a month. The markup is now the single largest avoidable line in your bill.

The break-even is not a hard number; it is the point where the monthly saving exceeds the small overhead of managing a provider key and a second invoice. For most teams that lands somewhere between ten and fifty thousand rows a month. Below it, managed mode is the rational choice. Above it, BYOK is.

When does managed mode win?

Managed mode wins whenever convenience outweighs a margin measured in dollars rather than hundreds of dollars.

• Low or occasional volume. A few hundred rows now and then saves pennies under BYOK. Not worth 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 run.
• One consolidated bill. Managed keeps AI, verify, and scrape on a single ReplyLabs balance instead of split across provider invoices.
• Active prompt design. While you are still iterating on a prompt, managed mode lets you test without provisioning keys. Move to BYOK once the workflow is stable.

A common path is to start managed, validate the workflow, then move the AI step to BYOK once monthly volume makes the margin visible. Because both modes use the same engine, switching is a key paste in Settings, not a migration. For the mechanics of running prompts across a column either way, see AI in Google Sheets.

Two things that do not change between modes

First, only succeeded rows are charged in either mode. Failed and skipped rows cost nothing on the ReplyLabs side regardless of which mode you pick, so a noisy input list does not inflate your bill.

Second, BYOK for AI is a Pro and Scale plan feature, and the seat fee applies either way. That is deliberate: ReplyLabs earns on the seat, not on a tax over your inference. The full mechanics of setting up a key, the supported providers, and the security model are covered on the bring your own API key page. For configuring a run in either mode, see the AI prompts help guide.

A quick decision rule

If you run more than roughly ten thousand AI rows a month and already have (or are happy to open) a provider account, use BYOK and pocket the markup. If you run less than that, or you want a single bill and zero key management, stay managed. New accounts get a $20 credit to try managed mode first, then move to BYOK once the volume justifies it. Start at the ReplyLabs home page to see the plans.

Common questions

Is BYOK always cheaper per row than managed?

Yes on the AI step. BYOK bills only the provider's raw rate, while managed adds a 1.25x markup and a $0.0025 base fee per succeeded row. The per-row saving is always positive; whether the total justifies a provider account depends on volume.

What is the break-even volume for BYOK?

There is no fixed number, but for most teams the monthly saving exceeds the overhead of managing a key somewhere between ten and fifty thousand rows a month. Below that, managed mode is usually the simpler choice.

Does managed mode ever cost less than BYOK?

Not on the per-row model cost, no. Managed mode is never cheaper per row, but it removes the cost of holding a provider account and consolidates billing, which can be worth more than the margin at low volume.

Do failed rows cost money in either mode?

No. Only succeeded rows are charged in both managed and BYOK mode, so failed or skipped rows are free on the ReplyLabs side.

Can I switch between BYOK and managed without losing work?

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.