The $1 government ChatGPT deal ends on 30 September — and the replacement has no spend cap, no minimum, and no budget line anyone has written yet

TL;DR: On 10 September 2026 GSA and OpenAI announced a 27-month OneGov agreement: 1 October 2026 → 31 December 2028. The $1-per-agency arrangement expires 30 September. What replaces it is metered token pricing at 50% off standard commercial rates, with the $15/user/month licence waived to $0, no platform-access fee, no minimum order, and no spend commitment. GPT-6 Astra is included — at 50% off that is roughly $5/$25 per million tokens, i.e. Opus 5 commercial rates. Eligibility widens from federal to state, local and tribal — about 23 million people, against 1M+ federal employees with access today. Daybreak Blue goes to every verified government entity at half price; Daybreak Red, the offensive tier, is available on request at full price. For buyers: a $1 line item cannot be overspent; a metered one with no floor, no cap and no commitment cannot be forecast — and neither GSA nor OpenAI has described a consumption control.

A dollar was never the price. It was the trial.

For the past year, a US federal agency could put ChatGPT in front of its staff for one dollar. More than a million government employees now have access. That was never a business; it was customer acquisition at national scale, and it worked.

On 10 September, GSA and OpenAI announced what comes next. From 1 October 2026 through 31 December 2028, agencies pay for what they consume, at 50% off OpenAI’s standard commercial rates. The $15 per-user monthly licence fee goes to zero. There is no platform-access fee, no minimum order, and no spend commitment.

Every clause of that is a genuine concession. Together they describe a bill that nobody currently knows the size of.

What the terms actually say

	Until 30 Sept 2026	From 1 Oct 2026
Structure	$1 per agency, per year	Metered, per token
Licence fee	—	$0 (was $15/user/month)
Discount	n/a	50% off commercial
Platform access fee	—	None
Minimum order	—	None
Spend commitment	—	None
Eligible	Federal	Federal, state, local, tribal (~23M people)
Term	—	27 months, to 31 Dec 2028

GSA Administrator Ed Forst framed it as positioning the federal government for AI-enabled capabilities. Laura Stanton, acting commissioner of the Federal Acquisition Service, called consumption-based access “the next logical step” now that agencies are integrating AI into regular operations. OpenAI’s Sam Altman described giving public servants secure access to the best tools.

Nobody in the announcement described a spend control. Neither did the trade coverage — Nextgov’s report notes the transition and quotes Stanton without any accompanying discussion of budgetary safeguards or limits, and FedScoop’s account contains no budget caps at all.

The arithmetic nobody has published

GPT-6 Astra lists at $10 input and $50 output per million tokens below the 272K context threshold. Halve those and a government tenant is at roughly $5 and $25 — the same rates Opus 5 charges commercially. That is a real discount on a real frontier model.

Now price a person. Assume a moderate user: twenty exchanges a day, about 2,000 input and 1,000 output tokens each.

Daily: 40,000 input + 20,000 output ≈ $0.20 + $0.50 = $0.70
Monthly: ≈$15 per user

That figure should look familiar. It is the licence fee that was just waived. The seat is free up to roughly the point where someone uses it lightly — which is an elegant piece of pricing, and a useful way to see the shape of the deal.

Then consider what agencies are actually being encouraged to do with it. Document analysis, case summarisation, and agent workflows run token counts ten to fifty times higher than a chat exchange. The same seat at 10x usage is $150/month; across a million employees that is a nine-figure annual line. At the frontier end, this desk’s own reading of OpenAI’s internal $600-per-researcher-day figure shows how far the top of the range goes.

These are illustrative figures on stated assumptions, not forecasts — and that is precisely the complaint. The real numbers do not exist publicly, and agencies have nineteen days to produce their own.

Why removing every friction is the risk

Each waived term is also a place where a budget conversation would have happened.

A platform fee forces a procurement decision.
A minimum order forces an estimate.
A spend commitment forces a ceiling — and ceilings are what become alerts.

Remove all three and adoption is frictionless in both directions. Nothing impedes a team from using the tool, and nothing impedes the bill, because no one ever had to write a number down to begin.

This is ordinary consumption-software strategy, not a scandal, and it is the same structure this desk described days ago in a commercial guise: a free harness that makes the spending decisions on a meter the vendor owns. The federal version is just bigger, slower to correct, and running against appropriations written before the meter existed.

The cyber clause is the bigger story

Buried under the pricing: every verified government entity is approved for Daybreak Blue, OpenAI’s advanced cyber-defender system, at 50% off. Government partners may also request Daybreak Red — the offensive tier — at standard commercial pricing.

Daybreak Red is the programme around GPT-5.6-Cyber, the model that completes ~95% of exploit-chain, privilege-escalation and authentication-bypass requests against 1.5% for the base model, and that produced a chainable zero-day in Chrome’s V8 engine. Access has run through a vetted partner tier of consultancies, major security vendors and specialist red teams.

What changes is the granularity of the gate. Approval now attaches to verified government entity status rather than to individually vetted organisations — and “government entity” now includes state, local and tribal bodies. This desk has argued that frontier cyber capability has become a clearance question rather than a purchase question, continuing the gated-access regime that became permanent in July. This agreement widens who holds the clearance by a large multiple, and it does so in a pricing press release.

The single-vendor question

One more thing worth noting for anyone building a government AI stack. The Defense Department’s own GenAI platform went deliberately multi-model, and the same department has demonstrated how quickly a vendor relationship can be vacated for reasons that have nothing to do with the technology. A 27-month single-vendor vehicle with zero switching friction at signup has considerable switching friction at month 20, once prompts, workflows and integrations have settled. The $0 licence buys adoption. It does not buy portability.

What to do with this

If you run a federal programme: instrument before 30 September. You have nineteen days to measure current usage and convert it to a projected token bill at $5/$25. That number goes in your FY27 planning, not a placeholder.
Set your own ceiling, because the vehicle has none. Per-agency and per-team spend alerts are your control now. No one else is providing one.
State, local and tribal buyers: assume no safety net. The absence of a minimum is what makes the deal reachable for you and what makes a surprise invoice possible. Cap it yourself at signup.
Separate the cyber decision from the pricing decision. Daybreak Blue eligibility arriving automatically with entity verification is a governance question that deserves its own review, not a checkbox inherited from a discount agreement.
Keep prompts and workflows portable. Twenty-seven months is long enough to build dependency and short enough to have to re-compete.
Non-government readers: run the same exercise this week. Whatever subsidised AI tier your organisation sits on, price it at commercial rates now. The transition from free to metered is the most predictable event in this market, and the only variable is whether you measured first. For teams comparing what that metering looks like across vendors, this week’s analysis of how an orchestrator bills the tokens you never see is the other half of the same lesson.

The ChatGPT review tracks OpenAI’s model tiers and commercial rates, and the best AI productivity tools rundown covers the deployment category most of these government seats fall into.

Frequently asked questions

What exactly changes on 1 October 2026?

The pricing model, not the access. The existing OneGov arrangement gives participating agencies ChatGPT for $1 per agency per year and expires on 30 September 2026. The replacement runs 1 October 2026 through 31 December 2028 — 27 months — and prices usage by token at 50% off OpenAI's standard commercial rates. The $15 per-user monthly licence fee drops to $0. There is no platform-access fee, no minimum order and no spend commitment, and access extends to OpenAI's current model line including GPT-6 Astra. GSA Administrator Ed Forst framed it as positioning the federal government for AI-enabled capabilities, and Laura Stanton, acting commissioner of the Federal Acquisition Service, described consumption-based access as the next logical step now that agencies are integrating AI into regular operations. The practical translation for a programme manager: the thing that cost one dollar in FY26 becomes a variable cost in FY27, and the variable has no ceiling written into the vehicle.

Is 50% off cheaper or more expensive than what agencies pay now?

Enormously more expensive, and that is not a criticism of the deal — it is arithmetic that someone needs to put in a budget. One dollar per agency per year is a rounding error; any metered bill at all exceeds it by orders of magnitude. The useful comparison is against the waived licence. GPT-6 Astra lists at $10 per million input tokens and $50 per million output below the 272K context threshold, so a 50% discount puts a government tenant at roughly $5 and $25 — which lands at Claude Opus 5's commercial rates. Now price a moderate user: twenty exchanges a day at about 2,000 input and 1,000 output tokens each works out near $0.70 a day, or roughly $15 a month. That is the same figure as the licence fee that was just waived, which is a neat way to see the shape of the deal — the seat is free precisely up to the point where someone uses it lightly. Move that same user onto document analysis or agent workflows, where token counts routinely run ten to fifty times higher, and the monthly figure per seat moves with it. These are illustrative numbers built on stated assumptions rather than forecasts, and that is rather the point: nobody has published the real ones.

Why does 'no minimum, no commitment, no platform fee' count as a risk?

Because each of those frictions is also a budget signal, and removing all three at once removes every natural moment where someone would have to forecast. A platform fee forces a procurement conversation. A minimum order forces an estimate. A spend commitment forces a ceiling, and ceilings are what turn into alerts. Strip them out and adoption becomes frictionless in both directions: nothing stops a team from using the tool, and nothing stops the bill from growing, because no one had to write a number down to start. This is the standard land-and-expand structure of consumption software and it is not unique to OpenAI or to government — it is the same shape as a free harness that decides how many tokens you spend, and the same shape the desk flagged in a private-sector guise this week. The federal version is simply larger, slower to correct, and running on appropriations that were set before the meter existed. Neither the GSA announcement nor the trade reporting around it describes a consumption control, a spend alert, or a per-agency cap.

What is the Daybreak Blue and Daybreak Red part of this?

The cyber tier, and it is the most consequential clause in the agreement. Every verified government entity is approved for Daybreak Blue — OpenAI's advanced cyber-defender system — at 50% off standard commercial pricing. Government partners may additionally request Daybreak Red, the offensive tier, at standard commercial pricing. Daybreak Red is the programme built around GPT-5.6-Cyber, a model that completes roughly 95% of exploit-chain, privilege-escalation and authentication-bypass requests against 1.5% for the standard model it derives from, and which produced a chainable Chrome V8 zero-day. Access to it has run through a vetted partner tier whose members are consultancies, major security vendors and specialist red teams. What changes here is the granularity of the gate: approval attaches to verified government entity status rather than to individually vetted organisations, and 'government entity' now includes state, local and tribal bodies. That is a meaningfully wider aperture on a capability this desk has described as a distribution decision rather than a training one.

Who becomes eligible that was not before?

State, local and tribal governments, for the first time under OneGov terms — OpenAI puts the expanded eligible population at roughly 23 million people, against the more than one million federal employees who have ChatGPT access today. Distribution runs through OpenAI directly, through resellers, and through cloud marketplaces. The scale change matters more than the discount. A federal agency has a CIO, a procurement office and an inspector general. A county government or a tribal authority typically has none of those, and is now being offered a metered frontier model with no minimum and no commitment through a reseller. The absence of a spend floor that makes the deal attractive to a small entity is the same absence that makes an unexpected invoice possible there, and small entities are precisely the ones least able to absorb one. If you work in or advise state or local government, the action item is to set your own ceiling, because the vehicle does not contain one.

If I am a non-government buyer, why should I care?

Because this is the clearest public example yet of the pricing transition every AI pilot eventually faces, and the terms are on the record. A vendor gives away the platform to build habit, the habit becomes infrastructure, and then the free tier expires and the meter starts — with a genuine discount attached, so the moment reads as a concession rather than a repricing. The structural lesson transfers directly: the danger is not the rate, it is that a year of usage accumulated with no cost signal attached, so nobody knows what the steady-state number is. Whatever free or heavily-subsidised AI tier your organisation is currently running on, the useful exercise this week is to instrument it as though you were already paying commercial rates, and find out what the bill would be before someone hands it to you. Agencies have until 30 September to do that. Most will not.

Sources

Related tool reviews

Questions or corrections? Email Pick Right. Want the full list? See all news.