Enterprise AI & Work
GSA's OpenAI Deal Trades $1 Access for Metered Usage
GSA's new OpenAI offer runs 27 months with a 50% usage discount. Agencies need a rate card and spending controls before October's transition.
Government AI buyers should prepare for metered spending after GSA announced a 27-month OpenAI agreement on September 10, expected to take effect October 1. Its 50% discount on token-based usage is a different commercial proposition from the earlier nominal $1-per-agency offer: access can become cheaper to provision while the budget becomes more sensitive to use.
The discount changes the unit, not the mission
GSA says the new agreement has no platform-access fee, minimum order, or spending commitment. It covers eligible ChatGPT models, including offerings in FedRAMP-authorized environments, with federal, state, local, and tribal eligibility and multiple ordering paths. Those terms remove barriers to starting small. They do not make unrestricted use free, nor do they establish that every model and deployment configuration an agency might want is covered.
The predecessor helps explain the transition. GSA’s August 2025 announcement offered each participating federal agency ChatGPT Enterprise for a nominal $1 for one year. That was an agency-level adoption offer, not a token rate. Comparing the new discount with that dollar as though they priced the same quantity would produce a dramatic but meaningless percentage.
One comparison is defensible. The original stated term was one year, or 12 months; the new announcement specifies 27 months. Combining those disclosures gives (27 − 12) ÷ 12 × 100 = 125% longer for the new stated offer term. That is a longer procurement planning horizon, not a 125% productivity gain, a spending cap, or proof that every underlying nominal token price is frozen. The offer’s duration and the agency’s future bill are different variables.
The missing denominator matters more than the discount. The September release does not publish a complete dollar-per-token rate card. An agency cannot take an unrelated public API price, halve it, and declare that its government contract will charge that amount. Ask which models qualify, how input, output, cached work, and additional features are metered, and whether different ordering channels add charges. Those are questions for the applicable order and commercial schedule, not gaps to fill with estimates.
The transition also needs an operational owner. GSA’s current Buy AI catalog still lists the original ChatGPT Enterprise offer through September 2026, while the new announcement describes an expected October start. Procurement teams should reconcile their existing order, actual expiry, authorized users, and the successor offer rather than assuming an announcement automatically modifies a deployed tenant. A continuous service experience and a continuous budget authorization are not the same thing.
GSA includes training and enablement resources in the new arrangement. That can support adoption, but the release does not quantify implementation labor saved for a particular agency. Budget for the agency’s own work: identifying approved tasks, reviewing data flows, training staff, setting cost controls, and measuring outcomes. The offer makes access easier to procure; it does not choose which administrative processes are safe or worthwhile to change.
A longer offer needs a shorter feedback loop
The practical beneficiaries are agencies with variable demand, small initial deployments, or uncertain adoption. No minimum spending commitment lets them test without committing to unused capacity. The strongest countercase is a heavily used workflow whose consumption is hard to predict. A percentage discount can coexist with a larger total bill when usage expands. Without an eligible rate card and measured consumption, neither the optimistic nor the alarming budget forecast is established.
Start with a defined service outcome rather than an account count. Select a task, record how it is completed now, and measure accepted output after introducing the tool. Keep review time and correction work in the calculation. A user who opens the product is not necessarily a user who saves time, and a completed model response is not necessarily completed agency work. Today’s lead on AWS evaluation sampling costs shows why the measurement layer itself also needs a budget.
GSA’s procurement guidance explicitly recommends small pilots, data-flow review, coordination with agency officials, and usage limits with regular consumption reporting. The security boundary remains agency-specific: its guidance tells buyers to consult their security team about authorization. A governmentwide commercial agreement does not by itself decide whether a particular dataset belongs in a particular environment. Obtain that decision before expanding access, not after the first useful demonstration.
The record also counsels against treating procurement headlines as blanket legal conclusions. In its December 2025 Ask Sage decision, GAO dismissed a challenge involving the original offer on contract-administration and interested-party grounds. The decision describes the offer as a price-list modification under a reseller’s existing schedule contract. It is not a finding that every future AI purchase, security configuration, or lock-in concern is resolved. The new order still deserves its own review.
That distinction extends our examination of Microsoft’s school AI contract scope. A broad institutional agreement can lower transaction costs without answering the deployment questions that make a product usable. Ask for export paths, administrative controls, an escalation contact, and the terms that apply when eligible models or features change. These are practical diligence requests, not claims that the present agreement omits every protection.
The evidence that would strengthen the adoption case is concrete: an eligible rate card, a clear ordering path, working usage reports, and pilot results showing a lower total cost for accepted work. The evidence that would weaken it is equally concrete: unpredictable metering, an unacceptable data boundary, or a review burden that overwhelms the benefit. Compare alternatives on those facts rather than their promotional entry prices.
Prepare the transition now; expand only against measured use. Contracting staff should confirm the successor order, finance should own consumption limits, and service owners should own outcome checks. A 27-month offer provides room to build something durable. It does not justify waiting 27 months to discover whether the meter and the mission are moving in the same direction.