Credits, not tokens: a pricing decision in the open

We had a long internal argument about pricing in February. Two camps: one wanted to pass through provider costs at a markup (so a user sees “5,000 GPT-5 input tokens · 2,000 output · $0.043” in their dashboard); the other wanted a single credit pool that abstracts the providers away.

We shipped credits. Here’s why, in the open, with the parts we found genuinely hard.

The problem with token pass-through

Token pricing makes perfect sense for a single-provider service. OpenAI’s own dashboard is fine. Anthropic’s own dashboard is fine. They’re each selling one product priced one way.

The moment you’re across providers, token pricing becomes intellectually honest and practically useless:

• Tokens aren’t comparable across models. A GPT-5 token and a Claude Opus token aren’t the same unit of work, and the conversion rate changes every few weeks as the labs reprice. Your bill at the end of the month is unpredictable in a way that single-provider users don’t experience.
• Multi-modal makes it worse. What’s an image “token”? An audio “token”? Every provider answers differently. Stitching it all into a unified dashboard means picking a normalization that’s wrong for at least half the providers.
• The bill discourages exploration. If switching from GPT-5 to Opus for one message changes your projected monthly cost by $4, you stop switching. That defeats the entire point of a multi-provider studio.

We watched ourselves stop trying alternate models during the early beta because we could see the price tick up. That’s the wrong incentive.

What we shipped

A monthly credit allowance per tier — 2,000 on Pro, 10,000 on Max. Credits are a shared pool. You spend them on chat, image, or video, whichever you need that month. The cost per generation is visible before you send: open the model picker and you see “Opus 4.7 thinking — 12 credits estimated for this turn.” Hit send, the cost confirms, you move on.

Two things this buys:

1. Predictable bills. Your cost is $20 or $100, full stop. You can’t accidentally rack up a $400 month because you happened to talk a lot to Veo.
2. No “should I switch models” tax. Switching is free, because the credit accounting is identical. Try Opus on the hard problem, drop to Haiku on the easy ones, generate a draft with Grok Imagine, finalize with Veo. The studio doesn’t punish you for using the right tool.

The objection we took seriously

The strongest argument against credits is cross-subsidization. If you use almost no credits and I burn through 9,000 of mine on Veo 4K clips, we paid the same $20. You’re subsidizing my video habit. That’s real.

Three things made us okay with this:

• The variance is bounded. With a 2,000-credit cap on Pro, a heavy user can’t burn 50x what a light user does. They can burn 1.5x, maybe 2x. That’s not a wild subsidy — it’s the same trade insurance and utilities have been making forever.
• Pricing transparency closes the gap. Because the per-generation cost is visible, users self-select toward the right tier. We see this in the data — heavy video users mostly land on Max within their first month. Light users stay on Pro and rarely hit their cap.
• Heavy users get a different conversation. If you’re consistently pushing past Max’s 10,000 credits, the right answer isn’t to make Pro more expensive — it’s to put you in touch with us about an enterprise pool tuned to your workload.

Where it doesn’t work

Credits would be the wrong call for a developer-platform business — inference-as-an-API where users are routing production traffic through us. Those users want per-token pass-through because they’re modeling unit economics, and the predictability we offer to consumers is the unpredictability they need to engineer away from.

We’re not that business. We’re a chat, image, and video product for people doing creative and operational work. Different shape, different right answer on pricing.

The footnote we wish we’d added sooner

The piece we underweighted in the original debate is how credit framing changes user behavior. People treat tokens like a meter (anxious to turn it off) and credits like a budget (willing to spend down to zero on things they care about). That’s not a coincidence — it’s a known pattern in consumer pricing, and it produced exactly the result we wanted: users explore the catalog instead of locking into one provider.

If you’ve been holding off on a model because you weren’t sure what it would cost — open the studio, hit it, look at the credit count. Worst case you burn 30 credits. The math is supposed to be boring now. That’s the point.