Satya Nadella wrote in July that with AI “you essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful.” I’ve already covered where his fix falls short. This post is about something narrower: what the second payment costs.

OpenAI has had the rate card online for almost two years. In December 2024 it announced a program that gives free daily tokens to organizations that agree to share their API inputs and outputs with OpenAI. The cap is up to 1M tokens a day across its larger models and up to 10M across the smaller ones, falling to 250K and 2.5M for usage tiers 1 and 2. In April 2025 OpenAI extended it and promised 30 days’ notice before ending it. Price trackers still list it as active.

None of that is new this month, which is what caught my attention. The price of the knowledge leaving your company has been public for 22 months, and almost nobody has set it next to the knowledge itself.

What the deal is worth

Let’s do the arithmetic. I couldn’t open OpenAI’s pricing page, so my yardstick is what price trackers list for GPT-5.4: $2.50 per million input tokens and $15 per million output tokens. Check the current rate before you quote this in a meeting.

One million tokens a day is worth between $2.50 and $15 daily. Over a 30-day month, $75 to $450. Over a year, $912 to $5,475.

That’s the going rate for handing over the inputs and outputs of your operation. If what travels is summaries of public articles, it’s free money. If what travels is client contracts, price lists, or the criteria your risk team uses to reject an account, $450 a month is a ridiculous price for something that, in Nadella’s words, “a competitor could never buy.”

The exposure sits in accounts you don’t control

Your company probably hasn’t switched the program on. Per OpenAI’s published policy, API data isn’t used for training unless the organization chooses it, and the switch lives on the data sharing settings page, which only organization owners can change.

The risk is in the accounts you don’t run. A contractor building your automation on their own OpenAI account. An agency testing with your real documents. A software vendor whose product calls the API underneath. Any of them may have switched the program on to cut test costs, and your data lands inside an agreement you never signed. That’s a possibility to rule out with an email, not something I’ve watched happen at a company.

One more detail from the fine print. In March 2026, forum users reported that GPT-5.4 mini billed them after 1M tokens instead of the advertised 10M. Support acknowledged that the documentation and the real model coverage might not match. Even the free credits need an audit.

What to check this week

Three things, in this order.

Your own OpenAI organizations come first. Open the data sharing settings on each one and confirm it’s off. If someone turned it on, find out why.

Next, your vendors. Ask in writing every vendor that calls a model API with your data whether any data-for-credits program is active on that account or on their subcontractors’ accounts.

Last, the contract. Add a short clause banning the trade of client data for credits, discounts or free tokens without written approval. It belongs next to the language that defines what counts as your intellectual property, which is what the contract has to protect in the first place.

Pricing your own data

If you want to use the program on purpose, there’s a tidy way to do it. Sort your traffic into three buckets: public or generic, internal but not distinctive, and what makes you different. Only the first bucket goes to an organization with sharing switched on. The other two go to one with the program off.

To sort, you need to know which data each process touches. That comes out of the Process Reality Map in AI Maestro, the map of how your company actually works. Without it, bucket three fills up with whatever someone pasted into a prompt on a Friday afternoon.

OpenAI put a price on your traffic: $75 to $450 a month per million tokens a day, at list rates. Before you accept or refuse, set your own price on what you’d be handing over.

Find out which data each process touches before you share anything