OpenAI tests outcome-based pricing after an 80% price cut

OpenAI CFO Sarah Friar spent her time at Goldman Sachs' Communacopia + Technology Conference on two numbers that matter more to your budget than any benchmark. The company cut the price of its low-cost Luna model by 80% and saw usage jump roughly tenfold. And it is experimenting with outcome-based pricing instead of charging per token. Those two facts point in opposite directions, and the gap between them is your planning window.

What actually happened

Speaking in San Francisco on September 8, Friar said enterprise revenue grew 32% from June to July against 20% growth in overall annualized revenue over the same stretch, and that the enterprise and consumer businesses reached a roughly even split by mid-year — ahead of OpenAI's own year-end target, according to Reuters.

The vertical push is the headline: chip design, life sciences, financial services. OpenAI used its own models to design its Jalapeno chip and reached tape-out in nine months. Friar's competitive claim was explicitly about price, not capability — deploying Luna against Z.ai's GLM 5.3 on a cloud layer, she said, "we are cheaper." Codex, meanwhile, has grown to 25 million users from roughly 100,000 at the start of the year.

Then the part nobody quoted in a headline: OpenAI is testing pricing tied to outcomes rather than usage, because buyers keep asking for a clearer return.

Why it matters for your business

Start with the 10x. AI pricing elasticity is now measured, not theoretical. An 80% cut produced roughly ten times the volume, which means the price cut was not generosity — it was a demand experiment that paid. The operator lesson is unglamorous and immediate: any automation you priced out in the last twelve months deserves a fresh spreadsheet. The document-extraction job that cost $400/month at old rates might cost $80 now. That's the difference between "someday" and "this quarter."

The outcome-pricing experiment is the one to watch warily. Per-token billing is ugly, but it is the only AI pricing model you can actually forecast: you know your call volume, you know your context size, you can cap it. Outcome pricing moves the meter to something the vendor defines — a resolved ticket, a completed task, a closed matter — and it prices your automation against the value it creates rather than the compute it burns. That is a fine deal when the automation is marginal and a terrible one when it works. The better it works, the more it costs.

So do two things while tokens are still metered and cheap. Re-run the math on the automations you shelved. And keep the model layer swappable, so that when a vendor changes how it charges, moving is a config edit rather than a negotiation you enter with no leverage.

Sources: Reuters, via Investing.com.