OpenAI’s Cheap Government Deal Will Be Judged by How Easy It Is to Leave

OpenAI plans to replace its $1-per-agency federal ChatGPT promotion with a government offer that charges according to use. The new agreement, announced September 10 with the General Services Administration, discounts eligible usage by 50%. A cheaper rate could help agencies expand useful work, but it does not put a ceiling on the bill.

The important change is the move from trying a product to budgeting for work performed through it. A nominal entry price makes an experiment easy to authorize. A continuing service becomes part of an organization’s operating decisions. For agencies, the strongest deal will be the one that combines useful work with a credible ability to change course. A large discount does not establish that second condition.

The current promotion expires September 30. The replacement is scheduled to run from October 1, 2026, through December 31, 2028, and extends eligibility to state, local and tribal governments. OpenAI says it will waive the standard $15 monthly license fee per user, with no minimum commitment.

Cheap access is a beginning

There is a straightforward public benefit to lowering the cost of experimentation. An agency should not have to make a large commitment merely to discover whether a tool helps with a real task. Waiving a license charge can reduce that initial hurdle. The case for the offer is strongest when the experiment produces enough information to support an informed decision about continued use.

The risk comes when successful experimentation becomes a reason to postpone that decision. Consider a hypothetical department whose staff start using the service to prepare recurring reports. Over time, templates, instructions and review habits could grow around it. None of that needs to be imposed by the supplier. People may build dependencies themselves because a tool is convenient, then discover that replacing it involves replacing part of their working routine.

Usage is measured in tokens, the small units of information an AI system reads or produces. OpenAI’s general Enterprise rate card distinguishes input, cached input and output. Cached input is material the service can reuse rather than process from scratch. The page tells customers to consult their agreement for applicable prices.

Usage pricing also changes what a manager needs to understand. The cost is no longer fully described by how many people have access. Two departments with the same number of accounts could use the service very differently. A comparison that treats every account as an equivalent unit would miss the work generating the bill. That makes a successful pilot an exercise in measuring demand as well as assessing the answers.

The public announcements and OneGov listing reviewed here do not identify a complete contract-specific rate card. Applying the advertised discount to an unrelated business price list would give false precision. An agency’s eventual bill also depends on what it asks the service to do and how often.

The discount is not the budget

This uncertainty is not evidence that the eventual bill will be excessive. It is a reason to avoid treating a percentage as a complete business case. A lower rate can make an expensive task affordable; it can also encourage much more use of an already affordable task. Those outcomes have different budget implications even if the advertised saving is identical. The relevant comparison is what the agency obtains for the expenditure.

Take a hypothetical department comparing a few short queries with repeated analysis of lengthy reports. Even with the same percentage discount, the amount of input and output can differ substantially. The useful budget question is the cost of completing that department’s work, not the discount in isolation.

In the report-writing example, a good result might be a shorter preparation cycle without a loss of accuracy. But the comparison would need to include checking and correcting the output, not just the time required to generate a draft. A low processing bill can coexist with an expensive review process. Conversely, meaningful staff time saved could justify a higher service bill. The purpose of measurement is to distinguish those cases.

GSA’s listing includes a one-year checkpoint to assess progress. It does not describe that checkpoint as a cancellation right. FedScoop separately reported procurement expert Jessica Tillipman’s warning that moving away from a deeply used system can disrupt an organization beyond the price of replacement software.

A checkpoint needs a real choice

The one-year checkpoint is potentially useful because it creates a moment to examine adoption before the agreement ends. Its value will depend on what an agency can do with the answer. A review that discovers weak results is more consequential if the organization can reduce its dependence without disrupting essential work. A meeting scheduled on the calendar cannot supply that flexibility by itself.

That is why the ability to leave deserves attention before the product becomes routine. In the hypothetical department, keeping the underlying documents usable elsewhere and understanding which instructions or processes depend on the service could preserve options. The point is not to demand that every vendor make itself interchangeable. It is to know what would need rebuilding, who would do it and how much disruption a switch would entail.

OpenAI is offering a lower barrier to adoption, and that can be worthwhile. The government’s side of the bargain should be disciplined learning: which tasks improve, what complete work costs and what alternatives remain practical. An introductory deal succeeds for the buyer when it produces better information and better choices. If it produces a dependency the buyer cannot evaluate or escape, the headline discount was never the most important price.