Federal AI Discounts Expire on September 30
On Monday, August 24, a federal acquisition team opening the public OneGov catalog will find three unusually cheap AI offers with the same end date. ChatGPT Enterprise is listed at $1 per agency. Gemini for Government is listed at $0.47 per agency. Claude’s public listing shows $1 per user. All three offers expire on September 30, 37 days from today.
The 37-day interval turns a promotional launch into a renewal decision. Procurement officials have to read the controlling terms, ask program offices which employees actually use the products, total the implementation work around each license, and decide whether to renew, narrow, compete, or exit. A usage report assembled after the offer ends cannot guide the order that must replace it.
Those headline prices made access easy to announce. Operating the products still required identity controls, decisions about permitted data, employee training, output review, approved workflows, contractor oversight, and incident preparation. Some offices bought direct employee access. Others tested a product through a limited cohort or a shared evaluation environment.
Two published counts show how far federal adoption has spread. OpenAI says more than one million U.S. government colleagues use ChatGPT Enterprise. The Government Accountability Office counted 282 reported generative AI use cases at 11 selected agencies in its 2024 inventory, up from 32 a year earlier. Neither count shows weekly use, the tasks that survived testing, or whether an agency received better service at lower total cost.
September 30 tests both price and evidence. A useful renewal file keeps five records separate: contracted access, provisioned users, active use, workflow results, and total operating cost. Collapse them into one adoption number, and a $1 offer can look successful before anyone follows the work from a prompt to a public result.
August 24 opens a 37-day renewal window
Open GSA’s current OneGov IT portal and the deadline is in plain sight. Claude, Gemini for Government, and ChatGPT Enterprise all carry September 30, 2026 expiration dates. Perplexity Enterprise Pro and Grok for Government appear beside them with March 2027 expirations. Federal AI purchasing continues; this first group reaches its price decision sooner.
September 30 is also the final day of the federal fiscal year. The teams collecting usage and quality records are working beside contracting staff closing fiscal 2026 actions and program offices defending fiscal 2027 needs. A temporary offer is crossing into base-budget territory on the busiest possible procurement date.
Public pages are a starting point, not an order file. Claude illustrates the difference. The current portal shows $1 per user, while GSA’s August 2025 launch release described $1 per agency. An authenticated term sheet and an agency’s order determine its obligations. A procurement team has to resolve that discrepancy before calculating a renewal baseline.
Gemini’s portal card shows $0.47 per agency, and ChatGPT’s shows $1 per agency. Those figures describe promotional access under OneGov. They do not disclose the price after September 30, usage commitments, optional services, migration work, or the terms an agency negotiated in a specific order.
GSA’s Buy AI page, updated on August 5, provides the more useful operating warning. It tells buyers to begin with the agency need, run a small pilot, review data flow and storage, involve technology, AI, data, security, and privacy officials, set usage limits, and review consumption. It also warns that AI costs can grow as use expands.
GSA’s instructions create the calendar. Program owners identify material workflows. Security and privacy teams check whether current use stayed inside approved boundaries. Finance gathers direct and indirect costs. Acquisition officials compare an extension with a competition or a narrower order. Employees receive notice and a workable alternative if access changes.
Thirty-seven days is enough for a disciplined review only if the evidence already exists. It is too short to reconstruct a year of use from anecdotes. Prompt counts without employee roles cannot show coverage. License assignments without login frequency cannot show use. Time-saving estimates without baselines cannot show net work. Vendor invoices without training, integration, review, and support cannot show total cost.
This is the budget handoff that catches many pilots. Temporary money pays for discovery and forgives incomplete economics. Recurring money needs an operating owner, a service level, and a reason this use ranks above another claim on the same appropriation. AI’s striking introductory price delays that conversation; it does not remove it.
Incomplete records do not prevent every decision. Mark the missing fields and reduce the scope. Renew a proven workflow instead of an entire eligible population. Extend one cohort while running a competition. Keep an evaluation environment while pausing an integration that never reached production. The deadline does not make weak evidence complete.
GSA priced access before agencies measured use
In April, GSA attached a large number to its new buying model. The agency said its first 20 unified agreements had produced discounts of up to 90% and $1.1 billion in first-year savings. OneGov used the government’s scale to reduce entry prices across widely used software and AI products.
That aggregate claim covers agreements involving Microsoft, Adobe, Google, ServiceNow, and other providers. It cannot be assigned to the three AI offers expiring in September. It also measures negotiated purchasing savings under GSA’s method, not the quality of a federal workflow or the value delivered to a member of the public.
OneGov’s public language changed as the program matured. In August 2025, Acting GSA Administrator Michael Rigas and Federal Acquisition Service Commissioner Josh Gruenbaum presented the OpenAI, Anthropic, and Google agreements as routes to rapid access, modernization, and adoption.
By April 2026, Administrator Edward C. Forst and Acting FAS Commissioner Laura Stanton were reporting aggregate savings and lower administrative burden. Both stages describe GSA’s program. Neither decides whether one agency’s next dollar belongs in one AI workflow.
The current public offer set looks like this:
| Public OneGov listing | Promotional public price | Public expiration | Renewal fact still needed |
|---|---|---|---|
| Claude for Government | $1 per user on the current portal; the 2025 launch release said $1 per agency | September 30, 2026 | Controlling order terms, eligible cohort, active use, service scope, next price |
| Gemini for Government | $0.47 per agency | September 30, 2026 | Product scope, active use, integrations, next price, exit process |
| ChatGPT Enterprise | $1 per agency | September 30, 2026 | Provisioned and active users, workflow outcomes, support scope, next price |
| Perplexity Enterprise Pro | $0.25 per agency | March 31, 2027 | Same evidence fields, with a later decision window |
| Grok for Government | $0.42 per organization | March 14, 2027 | Same evidence fields, with a later decision window |
Reading those five rows as a price comparison would hide most of the buying decision. Product scope, security authorization, data handling, accessibility, model behavior, administration, support, and contract rights differ. One product may fit research and fail an operational task. Another may perform well only after an agency pays for connectors, retrieval, evaluation, or professional services.
Two constraints survive the discount. Agencies still follow competition and fair-opportunity requirements when they place orders. Engineering, integration, cybersecurity, mission support, and other services retain their cost. Original equipment manufacturers continue to work through resellers, integrators, and small businesses that supply much of that labor.
For vendors, a near-zero first-year license can attract a large user population before the buyer knows which integrations are portable. Once an agency builds identity rules, document retrieval, evaluation sets, employee training, and support procedures around one product, the next price competes against switching costs created during the discount year.
Low introductory prices can also improve competition. They let multiple offices test credible products without consuming the full software budget. Teams can learn where general-purpose chat works, where a shared evaluation suite is sufficient, and where a mission workflow needs a specialized service. That learning has value even when an agency declines to renew.
A useful trial leaves assets behind. An evaluation set can test another provider. A task baseline survives a product change. A clear data inventory and an exit procedure reduce the next competition’s cost. Product-specific prompts and favorable employee stories create much less bargaining power.
OMB’s acquisition memorandum M-25-22 gives agencies a practical frame. It calls for performance-based acquisition, ongoing testing and monitoring, pricing transparency, protection of government data and intellectual property, portability and knowledge transfer, and rights that allow the government not to extend a contract. Each item belongs in the September file.
One million users is still a vendor numerator
OpenAI supplies the largest public reach claim. Its current government page says more than one million U.S. government colleagues are using ChatGPT Enterprise, alongside administrative tools, access controls, audit capabilities, and federal security authorizations. Federal use is no longer confined to a few innovation teams.
Yet the renewal denominator is absent. OpenAI does not publish the agency split, eligible population, provisioned seats, weekly or monthly active rate, task distribution, output quality, or independently reviewed outcomes behind the figure. “Using” could cover regular work, occasional access, a training exercise, or a single session during the measurement period.
An agency needs its own funnel:
- Eligible employees could receive access for an approved purpose.
- Provisioned employees received an account and working authentication.
- Trained employees completed enough instruction to use the tool and recognize restricted data or weak output.
- Active employees used it during a defined week or month.
- Workflow users completed a named task with the product.
- Outcome users improved time, quality, service, compliance, or another preselected measure after review work was included.
No stage substitutes for the next. Provisioning reveals technical reach. Activity shows whether the product entered work. Workflow completion shows task fit. An outcome comparison supports a renewal decision. A broad rollout may produce value in only two bounded workflows, which could still justify a smaller order.
A CIO can make a credible objection to this funnel. General access creates room for discovery, and forcing every employee to join a preapproved business case can suppress the unexpected uses that make a general-purpose tool valuable. Requiring project-level return on every research session would replace cheap experimentation with paperwork.
Keep an exploration account instead. Cap its population and consumption, provide training, protect employee feedback, and review the task mix at a fixed date. Operational workflows enter the stricter outcome file once an office asks for recurring budget, integration, sensitive data, or authority over a consequential decision. Discovery retains room to breathe without inheriting an automatic renewal.
GAO’s 2025 federal generative AI inventory offers a second adoption view. Eleven selected agencies reported 282 use cases for 2024. Of those, 84 were initiated, 75 were in acquisition or development, 46 were in implementation or assessment, 69 were in operation and maintenance, and eight had been retired.
Use-case counts differ from user counts. One operational system can serve thousands of people. A hundred experiments can involve a small central team. An inventory also reflects what agencies identify and report, not every informal use of an approved chat product. GAO’s snapshot should not be converted into a federal seat estimate or a success rate.
Its stage labels are useful for September. A buyer can connect every renewed license population to at least one workflow and name that workflow’s evidence state. An initiated experiment should not inherit the status of an operational service. A retired use case may represent good management if the agency learned that quality, cost, or risk failed the threshold.
Retirement evidence is particularly valuable during a promotional year. Teams face pressure to report adoption, and vendors benefit from large reach numbers. An agency that records why employees stopped using a tool can identify training failures, missing integrations, low trust, poor task fit, or a product that simply offered no advantage over an existing process.
Active use can fall for good reasons too. A benefits team may need a summarization workflow only during a policy cycle. An acquisition office may use a drafting assistant heavily before a solicitation and rarely after award. Frequency should follow the task’s cadence. A daily-use target imposed on episodic work rewards unnecessary prompts.
For that reason, the renewal denominator belongs at the workflow level. Name the population that performs the task, the period when the task occurs, and the alternative process. Then measure how many eligible people chose the product, how many completed the task, how often a reviewer corrected the output, and whether the result reached the next step.
Pennsylvania supplies a bounded public-sector baseline
Pennsylvania began with 175 employees, 14 agencies, and a year of feedback. That smaller state pilot supplies the clearest public workforce baseline, with limits visible enough to use.
Pennsylvania and Carnegie Mellon University studied 175 employees across 14 state agencies during a generative AI pilot. The March 2025 report received direct feedback from 136 participants. More than 85% described a positive experience, and participants reported an average of 95 minutes saved per day.
The 95-minute figure is attractive enough to overwhelm the rest of the report. Participants volunteered or entered through interested agencies, and the sample skewed toward employees aged 35 to 54 and people with postgraduate education. Time savings were self-reported. This was neither a randomized federal trial nor evidence of a causal annual dollar return.
Pennsylvania did move beyond the original cohort. In April 2026, Secretary of Administration Neil Weaver’s office reported more than 3,000 employees using approved generative AI tools across 35 agencies, with another 6,500 enrolled in required training. Those are state-reported reach and training figures, not an independent measure of task outcomes.
The expansion shows what workforce evidence can look like after a pilot. A labor-management group with employees represented by SEIU Local 668 meets on AI adoption. Pennsylvania’s unemployment chatbots match varied phrasing to preapproved responses and send out-of-scope questions to a live representative. That design names the employee voice, the answer boundary, and the human handoff instead of treating a chatbot session as complete service.
Its operational details are more portable than the headline. Nearly half of participants had never used ChatGPT before the pilot. Research, writing, and summarization became common uses. Some employees reported inaccuracy, uncertainty about permitted data, a steep learning curve, difficulty changing habits, or too little time to learn. Four percent of exit-survey respondents reported no use.
Imagine the renewal decision for a contract specialist who used a tool to produce a first draft of market-research notes. The model may shorten drafting, while the specialist still checks citations, removes unsupported claims, protects procurement-sensitive information, and aligns the file with agency rules. A timer that stops at the first draft counts gross model time. The workflow record continues through review and approval.
A program manager sees another part of the result. Faster drafting may clear a backlog, improve documentation, create time for supplier conversations, or reduce contractor hours. It can also raise expected output without changing staffing or service. The destination of released time determines who receives the benefit.
Keep Pennsylvania’s 95-minute estimate inside that boundary. A supported public-sector cohort perceived meaningful help. The sample and method also show why each federal agency needs a local baseline, observed workflow measures, and a record of review work before attaching dollars to time.
On August 11, the Census Bureau published a broader worker baseline. Its national survey story reported that 55% of workers had used AI for at least one of 11 job tasks at some point. Among those AI users, 24% used it daily in the previous week, 46% used it on some days, and 30% did not use it during that week.
Among recent users, 31% reported completing tasks one to two hours faster during the week. Another 25% reported less than one hour saved, 15% reported three to four hours, and 15% reported more than four. Ten percent reported no time saved, while 3% said AI added time.
Search assistance was the most common task at 37%, followed by writing and idea generation at 32% each, interpretation or summarization at 31%, and administrative work at 27%. These categories resemble common general-purpose AI uses. They still describe U.S. workers across sectors, not federal employees or one product, and the reported time is not a causal estimate.
Together, the Pennsylvania and Census evidence argues against two shortcuts. Low weekly frequency does not automatically signal failure, because many tasks are episodic. Broad access does not automatically signal value, because some users save no time or incur extra work. A renewal review needs both frequency and net task effect.
Employee experience belongs in that record. Training time should be paid and scheduled. Employees need a clear route for reporting inaccurate output or restricted-data uncertainty without turning the report into an individual performance mark. Cohort-level use can guide procurement while protecting workers from a prompt-by-prompt surveillance file.
Governor Josh Shapiro made the destination of time explicit in a March 2025 employee letter. His administration committed to employee and union input and said any saved time should be reinvested in customer service, customer outcomes, and public operations. A federal renewal file need not copy the policy, but it needs an equally clear answer about where released capacity goes.
Managers also need a workload account. If AI reduces first-draft time, reviewers may receive more drafts or more subtle errors. If a central team builds reusable prompts and evaluations, local teams may save work. If each office repeats the same setup, the cheap license can produce expensive duplication. Renewal evidence should show where work moved, not only where one task became faster.
Licenses end where integration invoices begin
License day one is the simple part. Mission work requires identity, approved content, records controls, security monitoring, accessibility support, evaluation, and incident processes. Employees need task instructions and a route to a human decision maker. Contractors or internal teams maintain those layers after the launch group moves on.
GSA’s OneGov guidance explicitly preserves this services market. It notes that agencies may still need engineering, integration, cybersecurity, and mission support, and that resellers, integrators, and small businesses remain part of delivery. The promotional license is one line in a wider cost file.
GAO’s 2026 review, Artificial Intelligence Acquisitions, examined 13 acquisitions and 44 contracts or agreements across the Departments of Defense, Homeland Security, Veterans Affairs, and GSA. The awards ran from September 2018 through February 2025, so the findings cannot rank today’s OneGov products.
Across those acquisitions, none of the four agencies had policy requiring systematic collection and application of lessons learned. GAO also observed that the quality of services surrounding an AI capability can matter more than the core model in an acquisition. Testing, data rights, performance monitoring, and the ability not to extend an arrangement all shape the result.
Service quality reaches employees directly. A reliable help channel can shorten the time between a weak answer and a corrected workflow. Poor integration can make an employee copy material between systems, repeat authentication, or abandon the approved tool for a faster unauthorized path. An inaccessible interface can exclude workers even when an agency reports broad account coverage.
Total cost should include at least six accounts: product access, cloud or consumption charges, integration and data work, security and evaluation, training and support, and employee review or correction time. Exit work forms a seventh account when an agency must export records, replace connectors, retrain employees, or rebuild an evaluation set.
An incumbent has earned an advantage when it performs well and the agency has learned to use it. That advantage becomes a procurement risk when the buyer cannot separate product value from avoidable switching friction. Portability tests reveal the difference. Can the agency export relevant records? Can it run the same evaluation against another product? Are prompts, retrieval pipelines, connectors, and workflow instructions usable elsewhere? Does a human process continue during migration?
Small businesses around a federal AI order deserve a clearer measure too. A renewal can sustain specialized integration, security, accessibility, and training work. A product switch can create migration demand while stranding product-specific expertise. Some service labor supplies the control and mission context that makes the product safe enough to use; treating all of it as removable overhead would erase that function.
Continuity is the strongest case for extension. Thirty-seven days may be too short for a fair competition, safe migration, and employee retraining. Disrupting a proven workflow solely because an introductory price ends would impose its own cost. A bounded bridge or narrower renewal can protect continuity while preserving the later competition.
A defensible bridge has dates and scope. It identifies the workflows that need continuity, the evidence they have, the competition milestone, the maximum period, and the data or integration work required for exit. An undefined extension converts deadline pressure into lock-in.
A September 30 renewal evidence file
Open a blank renewal spreadsheet and give each material workflow one row. Acquisition drafting, policy research, benefits correspondence, technical support, and document review remain separate. Products, contractors, and models appear as inputs. The program owner remains accountable for the result.
By September 15, each row should contain enough evidence for an acquisition, program, finance, security, privacy, data, workforce, and legal review. Empty fields stay visible. A missing baseline can support a limited extension with a measurement plan; it cannot quietly become a favorable result.
| File field | Evidence needed by September 15 | Decision on September 30 |
|---|---|---|
| Controlling offer terms | Authenticated term sheet, current order, covered product and services, unit, expiration, next-price options | Renew the correct scope, compete it, or let it end |
| User funnel | Eligible, provisioned, trained, monthly active, and workflow-active users by role and office | Narrow unused access or fund a defined cohort |
| Task mix | Named tasks, frequency, sensitivity, alternative process, and human decision point | Retain supported tasks and remove ornamental use |
| Baseline and outcome | Pre-use time, quality, error, backlog, service, or compliance measure; comparison period and limitation | Renew only where observed evidence meets a stated threshold |
| Human work | Training, prompting, checking, correction, escalation, mentoring, and work transferred between roles | Count net work and set staffing or support accordingly |
| Public or mission result | Closest result experienced by a claimant, taxpayer, operator, or mission team | Protect service quality and define a minimum floor |
| Total cost | License, consumption, cloud, contractor, integration, security, evaluation, support, employee time, and exit work | Compare full renewal cost with alternatives |
| Security and data | Approved data, access controls, records treatment, incident history, audit route, and prohibited uses | Close gaps, constrain use, or stop the workflow |
| Portability and exit | Export test, reusable evaluation, connector ownership, knowledge transfer, continuity process, and deletion obligations | Preserve bargaining power and a safe transition |
| Competition and decision | Market research, fair-opportunity record, alternatives, bridge rationale, owner, date, and approval | Renew, narrow, compete, or exit with a recorded basis |
Several denominators prevent one adoption rate from carrying the whole decision. Eligible employees reveal reach. Trained employees reveal whether access came with preparation. Monthly active users reveal recurring use. Workflow-active users connect use to a task. Outcome-qualified users connect the task to a result. Cost per outcome-qualified user is more demanding than cost per provisioned account.
Time measures need the same discipline. Record the old process from start to approved completion. Record the AI-assisted process through review, correction, and handoff. Keep quality or error beside elapsed time. Then document what happened to released capacity. A shorter draft that creates a longer legal review has moved work rather than removed it.
Quality needs a floor as well as an average. Agencies serve people with different languages, disabilities, documentation, and access to appeal. An average response score can improve while a smaller group receives unreliable help. Renewal evidence should include expected edge cases, a human escalation route, and a decision about which outputs may never act without review.
Cost belongs to the workflow’s full period. A first-year implementation charge may fall in year two. A discounted license may rise. Internal support may decline as employees learn, or grow as adoption expands. A comparison that uses the first year’s license and the second year’s labor will distort both options.
Competition records should preserve uncertainty. An agency may lack a comparable outcome for another product. It may know the incumbent’s migration cost more precisely than a challenger’s implementation cost. Present those ranges instead of forcing false precision. A staged competition can test the same workflow with the same evaluation and show where differences matter.
Four decisions are available on September 30:
- Renew when a material workflow has active use, a supported outcome, acceptable total cost, functioning controls, and contract rights that protect the agency.
- Narrow when value is concentrated in a defined cohort or task and broad access adds little.
- Compete when the need remains, alternatives are credible, and reusable evidence can support fair comparison.
- Exit when the workflow did not progress beyond access, outcome or control thresholds failed, total cost is excessive, or another process works better.
None of those decisions requires a universal verdict on a vendor. One agency can renew a product for technical research and exit it for correspondence. Another can keep general chat access while competing an integrated workflow. A third can use the government evaluation suite to postpone a large order until the task is ready.
Ownership keeps the file from becoming paperwork. The program official owns the workflow and service result. Acquisition owns the contractual path. The CIO and security teams own approved operation. Privacy and data officials own their controls. Finance validates total cost. Workforce leaders examine training, workload, and employee effects. One named executive resolves conflicts and signs the decision date.
Every evidence row should link to a dated source. A dashboard screenshot needs the query definition and time window. A time study needs the sample and comparison. A vendor claim needs its origin and limitation. An employee survey needs response rate and protection against retaliation. A savings figure needs the cost accounts it includes.
Agencies do not need to publish sensitive usage, security, or pricing details. They can still report the stage of evidence: access offered, cohort trained, workflow active, outcome measured, renewal decision made. That vocabulary would let taxpayers distinguish a product announcement from a public result without exposing operational data.
October starts with a second price
On October 1, an acquisition official will reopen an order whose memorable first-year price no longer answers the buying question. The relevant number is the cost of keeping a defined workflow running, including the people and systems around it. Some agencies may justify a much higher product price. Others may use the discount year’s knowledge to place a smaller order or make a clean exit.
OneGov’s low prices created a rare federal experiment across several frontier AI products. They reduced the cost of entry, increased vendor competition, and gave employees a chance to discover useful tasks. Those are real benefits even before an agency proves a financial return.
Promotional success can still hide a weak operating record. A million users does not identify one agency’s active cohort. An inventory of use cases does not show service quality. A self-reported hour saved does not show review work or the destination of released capacity. A $1.1 billion program-level savings claim does not reveal the result of one AI workflow.
A September file supplies the missing connection. It begins with the controlling order, follows an employee through a named task, includes the human checking and integration work, reaches a public or mission result, and ends with total cost and exit rights.
By September 30, a renewal official should be able to point to one employee population, one workflow, one measured result, one full cost, and one safe alternative. If the record cannot make that trip, the cheap license bought access and experience. It did not yet buy a basis for renewal.
Related Reading: