NextFin News - OpenAI is ending a pilot program that let US federal agencies use its AI models for just $1 per year, replacing the giveaway with usage-based pricing at half the standard rate, according to a statement from the General Services Administration. The change takes effect October 1 and runs for more than two years, closing out what the government's central purchasing arm called the steepest technology discount it had ever negotiated.
The Pilot Ends, the Meter Starts
The new arrangement moves federal workers onto a usage-based pricing system with a 50% discount from OpenAI's standard commercial prices, the GSA said. The agreement gives agencies access to OpenAI's full model lineup, including GPT-6 Astra, the company's newest frontier model that began rolling out September 3. OpenAI is also set to announce Thursday that the same discount will extend beyond the federal government to state, local, and tribal governments.
The pilot being retired dates to August 6, 2025, when OpenAI joined the GSA's OneGov marketplace to sell AI software at government scale. Under those terms, participating agencies could buy ChatGPT Enterprise for a nominal $1 per agency for one year. During that year, 3.5 million federal employees used ChatGPT and generated $1.4 billion in cost savings, the GSA said.
The combination is what makes the moment matter. A year ago, OpenAI was paying, in effect, for the privilege of putting its product in government hands. Now it is charging for it - and doing so from a position of strength, with a freshly launched flagship model and a rival, Anthropic, still untangling itself from a legal fight with the Pentagon.
The math of the transition is unforgiving. Under the pilot, the government's bill was fixed at $1 per agency no matter how heavily the tools were used. Under the new terms, the bill floats with consumption. That shift transfers the volume risk from vendor to buyer: if federal workers lean into Astra's agentic features and run far more queries, the savings that justified the pilot shrink even though the headline discount looks generous.
The Loss Leader Did Its Job
The $1 deal was never a pricing strategy. It was a customer-acquisition campaign dressed as a procurement contract. At a reported standard price of roughly $60 per user per month, with a 150-seat minimum and an annual commitment, a single midsize agency deployment would normally carry a floor near $108,000 a year. Charging $1 per agency meant OpenAI absorbed the full cost of compute, support, and deployment for millions of users.
Why absorb it? Frontier AI models face a procurement problem that consumer subscriptions do not. Government buyers cannot simply sign up on a website. They need pre-negotiated contract vehicles, security reviews, and approved purchasing pathways. OpenAI cleared that friction by accepting a nominal fee through the GSA's Multiple Award Schedule - the same catalog the government uses to buy everything from office supplies to cloud services. It bought distribution.
Now the distribution is built. With 3.5 million federal employees already inside ChatGPT Enterprise, the switching costs have flipped. They sit on the government's side of the table, not the vendor's. The move from a fixed $1 fee to usage-based pricing is the tell: OpenAI no longer needs to subsidize access because the habit is formed.
"One of the best ways to make sure AI works for everyone is to put it in the hands of the people serving the country," OpenAI CEO Sam Altman said when the original partnership was announced.
That sentiment has not changed. The pricing has.
There is a second reason the loss leader made sense, and it is financial rather than strategic. OpenAI's infrastructure costs rise with every token generated. A free or near-free government rollout let the company train its operational muscle - support workflows, deployment playbooks, security review processes - on a large, demanding customer without the margin pressure of a paying contract. By the time the bill arrives, the service machine is already tooled. That is a classic loss-leader sequence: subsidize adoption, harden operations, then monetize the installed base.
Why the Timing Works
Three things line up. First, the new deal takes effect October 1 - the start of the federal fiscal year, when agencies refresh budgets and contract authorities. Second, it arrives one week after the September 3 launch of GPT-6 Astra, giving agencies a commercial reason to upgrade rather than a purely financial reason to stay. Third, the competitive field has cleared.
Anthropic negotiated the same $1-per-year price days after OpenAI's 2025 deal and staged a high-profile Washington event with CEO Dario Amodei. But that government push was quickly overshadowed by a fight with the Pentagon over surveillance and autonomous-strike guardrails. Defense Secretary Pete Hegseth labeled Anthropic a supply-chain risk, and President Donald Trump ordered agencies to stop using its tools. A federal judge in California issued a preliminary injunction in March, and last month ruled that the administration must lift the ban - though the Pentagon can still bar the technology from its own work, and a separate legal argument awaits a three-judge panel.
OpenAI, by contrast, announced on February 28 that it had reached an agreement to supply the Pentagon, with safeguards against domestic surveillance and no use for fully autonomous weapons. The asymmetry is stark: one vendor is litigating its way back into government trust while the other is raising prices.
The guardrail question deserves its own line, because it is the quiet precondition for everything else. ChatGPT Enterprise does not use business inputs or outputs to train OpenAI's models, a provision that carried over into the government deal. For agencies handling sensitive data, that data-handling promise is worth more than a price discount. OpenAI's willingness to lock it in - and Anthropic's parallel insistence on the same protection in its own negotiations - signals that data sovereignty, not dollars, is the real gatekeeper in government AI sales.
The Discount Is the New Anchor
The first-order story is simple: the government pays more. The second-order story is that the 50% discount becomes the reference price for the entire public-sector AI market - and eventually for the private sector too.
Here is the mechanism. OpenAI does not publish a sticker price for ChatGPT Enterprise; contracts are negotiated. Reported 2026 contracts cluster around $60 per user per month, with a 150-seat minimum and annual prepayment, implying an entry cost near $108,000 a year. A 50% government discount effectively sets a public reference point at roughly $30 per user per month. State, local, and tribal governments - the next tranche OpenAI will announce - will benchmark their negotiations against that number. So will private buyers. Once a discount becomes the market's reference price, the list price becomes decorative.
That is the trap built into deep public-sector discounts: they scale. The federal government is the largest single technology buyer in the world, and its contract terms ripple outward. A 50% discount that looks generous on a per-seat basis can compress what every downstream buyer is prepared to pay. Vendors hate this dynamic, which is why most resist publishing government rates. OpenAI has not published the rate either - the 50% figure comes from the GSA statement - but the number is now in the market, and reference prices are hard to un-ring.
There is also the usage-based element, and it matters more than the headline discount. A flat $1 fee capped OpenAI's revenue regardless of how much the government consumed. Usage-based pricing aligns revenue with consumption - and with the arrival of agentic models like Astra that can run multi-step workflows across software, consumption per user can rise sharply without any change in the posted rate. The government's $1.4 billion in reported cost savings from the pilot year may be followed by a much larger bill, not because the price per unit rose, but because the units consumed did.
Consider the arithmetic. If even a fraction of the 3.5 million federal users who tried ChatGPT during the pilot convert to regular paid usage, the addressable government spend moves from a nominal $1 per agency into the tens or hundreds of millions annually, depending on seat count and consumption intensity. At the reported $60-per-seat list price, outfitting the full 3.5 million-user base would imply a list-value exposure well above $2 billion a year; at the 50% government rate, that halves to roughly $1 billion. The actual figure will be far lower - not every user converts, and consumption varies - but the order of magnitude shows why the pilot was never going to last.
Cyclical Promotion, Structural Pricing
This is a structural shift, not a cyclical price adjustment. A cyclical move would be a temporary discount expiring and snapping back to list price. What OpenAI is doing is different: it is converting a promotional distribution channel into a permanent revenue line with a new pricing architecture, and extending it across government tiers.
The evidence is in the contract's shape. The deal runs more than two years. It covers the full model roadmap, including models not yet released. It expands from federal to state, local, and tribal governments. None of that is consistent with a promotion ending. The $1 pilot was cyclical - a mean-reverting loss leader that served its purpose once adoption hit 3.5 million users. The usage-based framework is structural - it will not revert on its own because it is now the procurement baseline.
The risk to that call is commoditization. If frontier models become interchangeable and price becomes the only differentiator, the 50% discount could keep falling toward the marginal cost of inference. That bear case is not trivial: industry estimates put Anthropic at roughly 40% of enterprise LLM API spend versus OpenAI's 27%, despite Anthropic's far smaller consumer reach. But for the federal government specifically, the switching costs, security reviews, and procurement integration already completed make near-term churn unlikely.
There is a parallel to watch in cloud computing. When hyperscalers first courted government work, they accepted deeply discounted enterprise license agreements to get foot-in-the-door. Once workloads migrated and re-architecting costs mounted, pricing normalized and consumption became the growth engine. OpenAI is running the same playbook, one generation later, with models instead of virtual machines.
The Counter-Thesis
The strongest case against this read is that OpenAI is leaving money on the table by discounting at all. A company running an annualized revenue rate near $40 billion and valued at roughly $852 billion does not need to give the government half off. Microsoft's fiscal 2026 filing with the SEC disclosed $24.1 billion in commercial arrangements with OpenAI for the year ended June 30 - meaning a single vendor relationship absorbed about as much revenue as OpenAI's entire annualized run rate. In that light, a 50% government discount looks less like pricing power and more like a company still buying legitimacy ahead of a public listing. OpenAI confidentially filed its S-1 in June, targeting a September IPO at a valuation between $852 billion and $1 trillion.
The counter-argument has teeth. Public-market investors scrutinize discount structures, and a high-profile half-price deal with the federal government invites questions about whether reported revenue reflects real willingness to pay. If the IPO prospectus shows government revenue propped up by a discount that would not survive competitive bidding, the multiple could compress.
But the counter-thesis mistakes the discount for the point. The point is the pricing architecture. A 50% discount on a usage-based contract with no consumption cap exposes OpenAI to uncapped upside if government usage grows, whereas the $1 deal capped revenue at $1 no matter how much the government consumed. Legitimacy before an IPO is a bonus; the structural change is that revenue now scales with adoption instead of being capped by a promotional fee. The discount is the hook; the meter is the business.
The signal that would prove this view wrong is concrete. If, within two contract quarters, OpenAI reverts to flat per-seat pricing for government customers or deepens the discount beyond 50% for new federal deployments, the structural-pricing-power thesis fails and the giveaway was simply extended. Watch the GSA's Multiple Award Schedule catalog for the listed rates after October 1.
What Comes Next
Who benefits and who is exposed. OpenAI benefits first: it converts a loss-leading pilot into a multi-year revenue stream with upside tied to consumption. Federal agencies benefit from continued access to frontier models with guardrails that keep government data out of model training - a provision that carried over from the pilot. The exposed parties are taxpayers, who will now foot a real bill for technology they got nearly free for a year, and Anthropic, which is fighting for government trust while its rival locks in terms.
By time horizon:
- Short term (next quarter): budget offices absorb the new pricing into fiscal 2027 plans; usage-based bills will be lumpy as agencies test Astra's agentic features and discover where consumption spikes.
- Medium term (two to four quarters): the 50% reference price spreads to state and local governments, compressing what those buyers were prepared to pay; Anthropic's legal position clarifies, restoring a second supplier for agencies that want one.
- Long term (beyond a year): the question is whether usage-based pricing proves durable or whether political pressure - and the optics of a roughly $40-billion-revenue company discounting to government - forces a return to flat-rate government pricing.
Scenarios. Base case: the usage-based model holds, government AI spend rises with consumption, and OpenAI books a steady public-sector revenue line that diversifies it away from dependence on its largest commercial partner. Upside case: Astra's agentic capabilities drive a step-up in per-user consumption, the state and local expansion opens a second growth leg, and the IPO prices at the top of the targeted range. Downside case: political scrutiny of government AI discounting intensifies ahead of the listing, Anthropic's legal victories reopen competition and push discounts deeper, or consumption fails to ramp and the meter never produces the revenue the architecture promises.
OpenAI did not just raise the government's price - it changed the meter. The $1 deal charged nothing for everything the government used; the new deal charges for every unit consumed. That is not the end of a promotion. It is the moment a loss leader becomes a business.
Data as of September 10, 2026.
Explore more exclusive insights at nextfin.ai.
