OpenAI cut government licence fees to zero. The usage meter still runs

The offer is large and precisely bounded

OpenAI and the US General Services Administration have announced a 27-month agreement covering eligible federal, state, local and tribal government organisations. From 1 October 2026 through 31 December 2028, the normal $15 per-user monthly licence fee falls to zero, there is no minimum commitment, and eligible usage receives a 50% discount.

The scale in the announcement needs careful language. OpenAI says more than one million government employees already have access through existing arrangements. The expanded offer makes a wider public-sector workforce of approximately 23 million eligible. It does not say 23 million people have enrolled, actively use the service or send paid requests. Eligibility is the size of the addressable population, not the installed base.

The agreement also gives verified government entities access to Daybreak Blue, OpenAI's defensive cyber capability, at half standard commercial pricing. Organisations can request Daybreak Red for advanced vulnerability research, exploit validation and red teaming at standard pricing. Buyer guidance, onboarding, spending controls, training and FinOps support are included in the programme.

The phrase $0 access is therefore accurate for one line of the bill. The meter remains on for consumption and specialist cyber access.

The commercial terms published by OpenAI on 10 September 2026.

Item	Published term	What it does not establish
User licence	$0 instead of $15 per user per month	No guarantee of unlimited model use
Usage	50% off eligible usage costs	No fixed total bill without volumes and workload mix
Daybreak Blue	50% off standard commercial pricing	It is discounted, not included at zero cost
Daybreak Red	Available by request at standard pricing	Access is not automatic
Term	27 months	Future pricing after 31 December 2028
Population	About 23 million eligible workers	Active users or adoption rate

A free seat can still create a large variable bill

The arrangement removes a predictable fixed cost and discounts a variable one. That changes procurement behaviour.

Consider a hypothetical agency enabling 10,000 staff. At the published list fee of $15 per user each month, the removed licence line is worth $1.8 million a year. That is a real saving. If only 2,000 staff use the service regularly, the avoided licence cost still applies to all enabled seats while usage accumulates only where work occurs. If automation later turns one analyst's task into thousands of model calls, the variable line can grow without adding another user.

This is why cost per seat is the wrong operating metric for agentic and API-based use. An employee asking for a summary, a developer running repeated code agents and a cyber team validating vulnerabilities consume very different resources under the same organisational agreement.

A public buyer should track at least four numbers: eligible people, enabled accounts, monthly active users and cost per completed workflow. The first is a market statistic. The last three determine the service and its budget.

Why OpenAI is willing to remove the licence fee

The commercial logic is straightforward. Removing the entry fee reduces the friction of putting the product in front of public employees. Usage revenue remains, specialist capabilities remain priced, and a 27-month term gives agencies time to build processes, templates, integrations and training around one platform.

Once a model becomes part of a tax workflow, public-health review or security operation, switching involves more than moving accounts. Prompts, evaluations, connectors, records schedules, staff skills and assurance evidence accumulate around the service. That is normal platform economics rather than proof of improper lock-in. It is still something procurement should measure before the first integration is built.

The examples OpenAI publishes illustrate the adoption case. It says a CDC literature-review pilot produced most initial reports in under 30 minutes and 92% of participating experts reported productivity gains. Georgia's Department of Revenue reduced a form-digitisation task from as long as two weeks to 15 minutes. Those are provider-reported examples, not a common benchmark across all government work. Agencies should preserve the local baseline so they can demonstrate whether the result survives scale.

The cyber offer raises the assurance bar

Defensive cyber access is a material part of the agreement. Daybreak Blue is intended to help defenders find vulnerabilities, analyse malware and build security tooling. Daybreak Red moves closer to exploit validation and advanced red-team activity.

That capability can improve public-sector defence and concentrates risk in credentials, environments and output handling. A model permitted to inspect malware or test an exploit should not share the same identity, network reach or approval policy as a general office assistant. "Government access" is not one trust level.

OpenAI says participating services retain their applicable protections and that ChatGPT Enterprise does not use business inputs or outputs to improve its models. Public organisations still need to decide which records may enter the service, where API integrations store outputs, which administrators can retrieve conversations, how legal holds work, and what evidence survives an incident.

FedRAMP Moderate availability addresses an important federal security baseline for eligible OpenAI services. It does not replace system authorisation, data classification, use-case review or local access control. Compliance describes the boundary assessed; it does not decide whether a particular workload belongs inside it.

Controls to establish before broad enablement turns into embedded dependency.

Decision	Minimum evidence
Who may activate accounts	Named owner, approved population and deprovisioning feed
Which data may be entered	Classification rule, examples and enforcement point
Who may create API projects	Separate credentials, spending limits and workload owner
Who may use cyber models	Role approval, isolated environment and authorised target scope
Whether the service delivered value	Pre-AI baseline, quality review, time saved and total usage cost
How the agency can leave	Export format, deletion evidence, connector inventory and replacement test

The P.K. view

This is a significant public-sector distribution agreement and a useful lesson in reading AI pricing. The headline price applies to identity. The operating cost follows activity.

The deal can be excellent value if agencies use it to remove repetitive work and preserve evidence of quality. It can also hide waste if free account creation is mistaken for successful adoption and discounted consumption is treated as unlimited. Fifty per cent of an unmanaged variable is still unmanaged.

The first dashboard should therefore show enabled accounts, active users, usage by workload, accepted outputs and human review time together. A service that saves one department thousands of hours may justify substantial consumption. A service that creates millions of summaries nobody uses is expensive at any discount.

The procurement question is not whether the licence is free. It is what the organisation will pay per verified public outcome, which data and authority the service receives on the way, and whether it can still answer those questions when the introductory term ends.