2026-05-16
Why we stopped asking customers for an Anthropic API key
We launched Replai on a BYOK model — bring your own API key, we don't mark up your tokens. It was honest. It also gated our actual customers out of the funnel. Here's what we changed and why.
When we launched Replai, every paid plan asked you to bring your own Anthropic API key. The pitch was clean — we don't mark up your tokens. You pay Anthropic directly, at cost.
It was honest. It was also the wrong default.
The problem we missed
Replai's actual customer is a SEA small-business owner. A tuition centre operator. A retail brand running a Hari Raya promo. An e-commerce founder who can't keep up with WhatsApp DMs.
None of those people have an Anthropic API key. None of them want to open a console.anthropic.com account, paste in a credit card, generate a key, copy it into our settings, then check an AWS-style billing dashboard every month.
We told them the price was $19/month. The actual time-to-first-reply was: sign up → hit a "paste your API key" wall → realise they need a separate Anthropic account → realise they need a credit card → close the tab.
We were selling to one audience and shipping for another.
What we changed
Replai now covers the AI cost on every paid plan. One price, includes the AI.
This isn't a cosmetic change. The numbers had to work first. Three things let them:
- Defined a "conversation" as a billing unit. Up to 5 bot replies in a thread. A longer back-and-forth counts as more than one. This makes the cost ceiling per customer predictable instead of unbounded.
- Tightened the caps. Starter is 120 conversations a cycle, Growth is 300, Pro is 600. Sized so LLM spend stays around 20% of net revenue at typical usage.
- Added a Conversation Pack overage. +100 conversations for $9, anytime. A predictable escape valve for the campaign that pushes a customer over their cap. No surprise bills.
The combination is cleaner than BYOK ever was. The customer has one price to think about. We absorb the AI-cost variance.
What it costs us
A typical Replai conversation runs on Claude Haiku and costs us a few cents, including the safety-net classifier that runs on every reply. At a 120-conversation Starter cap, the AI cost stays well under the gross-margin floor we set.
The pricing comment in our source code reads "caps sized to keep LLM spend near 20% of net revenue." That's the design target. It's intentionally conservative — we'd rather under-promise margin than under-deliver service.
The Pro plan used to say unlimited conversations. That's gone. Uncapped usage + Replai-funded AI is a blank cheque, and the only customers who'd notice it are the ones who'd cost us the most money. Pro now has a real number (600/cycle), and a true power user is routed to the Managed tier where pricing is scoped to actual volume.
What we'd say to anyone else building a similar product
If your customer is non-technical, BYOK isn't cost transparency. It's a wall. The technical buyer who likes BYOK is the secondary audience for most SMB AI products — not the primary.
The right place for BYOK isn't pricing. It's a power-user feature for the slim slice of customers who genuinely want to provision their own keys, usually because they have enterprise procurement or compliance constraints. Build it as an optional override, not as the default. Most of your audience just wants to pay one number and have it work.
We learned that the slow way. Hopefully the next team building in this space can skip the lesson.