A Frontier Model Launch Is a Repricing Event: A Six-Step Checklist for Founders

Key takeaways

• OpenAI listed GPT-6 Astra at $10 per million input tokens and $50 per million output tokens, with a fast mode at $20/$100 for up to 2x the speed.
• A separate speed tier at double the price means latency is now a purchasable input, and most SaaS pricing pages have nowhere to put that cost.
• Astra usage sits inside existing ChatGPT subscription allowances, with extra credits purchasable on top. That hybrid shape is quietly becoming the default AI pricing model.
• Your exposure is not the rate card, it is COGS per task multiplied by tasks per user per month.
• Run the checklist below before you decide whether to adopt, not after.

Why a model launch is a pricing event

Most teams read a launch post looking for capabilities. The line that actually reaches your P&L is further down the page: the rate card, the tiering, and whether usage is included or metered.

For GPT-6 Astra, OpenAI published standard API pricing of $10 per million input tokens and $50 per million output tokens. It also shipped a fast mode at twice the rate for up to twice the speed, and it noted that Astra usage falls inside existing subscription allowances, with additional credits available for purchase.

Each of those three facts is a pricing decision you now have to make in your own product.

What does a separate speed tier do to your pricing page?

This is the part worth sitting with. Fast mode is not a better model, it is the same model delivered sooner, at 2x the price.

That converts latency from an engineering property into a line item. If your product promises a response inside a certain window, you are no longer choosing between models, you are choosing how much to pay for time. And if you sell one flat price to everyone, you absorb that entirely.

Providers have been doing this for a while with priority and batch tiers. The difference now is the size of the spread. Doubling your inference COGS to hold a latency promise is not a rounding error on a 70% gross margin, it is the difference between a healthy tier and a loss-making one.

The practical response is not complicated: if speed costs you double, speed should be something a customer can choose. Either a faster tier, a per-request option, or an SLA that is priced rather than assumed.

The hybrid shape is becoming standard

Notice how Astra is sold to end users: included within existing plan allowances, with credits purchasable for more. That is neither pure subscription nor pure usage-based. It is a subscription with a metered overflow.

That shape exists because it solves the thing flat pricing cannot, which is a small number of heavy users destroying the margin on everyone else. It is worth copying, and it is worth modelling before you copy it, because the allowance size is the whole design. Set it too high and your power users are free-riding. Set it too low and every customer hits a wall in week two and churns.

The six-step repricing checklist

1. 1Re-measure COGS per task on the new model. Not per request. Per completed unit of customer-visible work, retries and tool calls included.
2. 2Recompute margin per tier. COGS per task times average tasks per user per month, divided into what that tier pays. Do this for each tier separately, because the answer usually differs sharply.
3. 3Find your break-even usage. At what monthly task count does a given tier stop making money? Then check what share of your customers is already past it.
4. 4Decide whether latency is a product. If fast mode is on the table, price it as an option rather than eating it silently.
5. 5Size the allowance. If you go hybrid, pick the included volume so it covers a clear majority of customers, and set the overage rate so heavy usage stays profitable rather than merely tolerable.
6. 6Set a re-check trigger. These launches now arrive every few months. Decide in advance what rate change forces a review, so you do not find out from a bank balance.

A worked example, illustrative only: say your average customer runs 400 tasks a month at 5,000 input and 2,000 output tokens each. At $10/$50 that is roughly $0.15 per task, about $60 a month in inference. On a $99 plan you have around 39% gross margin before every other cost. On a $199 plan you have 70%. Same product, same model, and only one of those tiers survives a price increase.

What to do if the numbers do not work

You have four levers, in rough order of how quickly you can pull them: route cheaper models to cheaper tasks, cut tokens per task through prompt and context discipline, cap or meter the heaviest usage, and finally raise price. Most teams reach for the fourth first because it is the only one that does not require instrumentation. That is backwards.

Takeaway: treat every frontier launch as a scheduled margin review, because your provider changes your cost structure without asking and your pricing page will not update itself.

If you want to see what a rate change does to each of your tiers before you commit, you can model tiers, margins and break-even in Calcaas.