If your agency runs Claude on the $1-per-user OneGov deal, that price ends Friday, and from November 1 you pay for tokens at 40% below Anthropic's commercial rates, with no cap unless you set one. Anthropic announced the expanded GSA agreement on October 27: a 40% discount on Claude Enterprise, Claude for Government and the Claude API, no minimum spend, and for the first time eligibility for state, local, tribal and territorial governments. GSA's OneGov listing marks the $1 offer as expiring October 31, 2026 and lists the new consumption offer through January 31, 2029.
A state CIO gets a door that was closed. A federal CIO gets a bill that used to round to zero and now moves with how much people use the tool. Both now have to price Claude against OpenAI, which made the same move seven weeks earlier with a bigger discount.
What Changes for Federal Agencies on November 1?
Federal agencies move from a flat per-user price to metered billing on November 1. The original deal, announced August 12, 2025, put Claude for Enterprise and Claude for Government in front of all three branches for $1, with the contract set to expire after one year. GSA's OneGov listing then extended it by a month in September, moving the end date from September 30 to October 31.
That extension has now run out. The replacement offer on GSA's listing reads "40% discount on tokens for consumption-based access to all Claude models available to government," with $0 commitment and $0 minimum orders, sold through Anthropic directly, resellers and supported cloud marketplaces, and bought on the Multiple Award Schedule (GSA OneGov). Anthropic says MAS ordering opened on the day of the announcement, and that the discount reaches cloud-provider purchases "in the coming months" (Anthropic).
Consumption pricing means you pay for the tokens your users and applications actually process, with no seat fee and no commitment. That cuts both ways. No one forces you to buy capacity you will not use, and no one stops a pilot that goes viral inside an agency from running up a bill.
The authorization picture carries over. Per Anthropic, Claude for Government has been generally available since September 30 at FedRAMP High, Claude is FedRAMP High authorized on Amazon Bedrock and Google Vertex AI, select models are authorized for DoD IL4 and IL5 in AWS GovCloud, and Claude Enterprise is in process for FedRAMP Moderate through FedRAMP 20x (Anthropic). We covered the records and billing controls in Claude for Government when it launched: Claude for Government Keeps the Only Copy of Each Chat on a Laptop.
What Does 40% Off Actually Cost?
At Anthropic's current list prices, 40% off puts Sonnet 5.5 at $1.20 per million input tokens and $6 per million output tokens. The math, from Claude's pricing page:
| Model | List input / output (per 1M tokens) | At 40% off |
|---|---|---|
| Opus 5.5 | $4 / $20 | $2.40 / $12 |
| Sonnet 5.5 | $2 / $10 | $1.20 / $6 |
| Haiku 5.5 (prompts up to 100K tokens) | $0.10 / $0.50 | $0.06 / $0.30 |
Run one assumption through it. Say a typical analyst sends 1 million input tokens and gets 200,000 output tokens back each month on Sonnet 5.5. That is $4.00 a month at list and $2.40 at the OneGov rate. A developer running Claude Code agents can burn through many times that, and long-prompt Haiku traffic is billed at a higher tier ($0.50 input above 100K tokens, per the same page). Your number depends entirely on who uses it and for what, which is why the first metered month is the one to watch.
One open question you should put to Anthropic or your reseller in writing: does the 40% touch the Claude Enterprise seat? Commercially, Claude Enterprise costs $20 per seat per month, billed annually with a 20-seat minimum, with usage billed on top at API rates. Anthropic's post says the discount covers Claude Enterprise; GSA's listing describes it as a discount "on tokens," and The Next Web's account describes it as covering token-based usage across those products. Neither page says what happens to the seat fee or the 20-seat floor under the OneGov terms. A small county with eight users has a very different bill depending on the answer.
How Does Anthropic's Offer Compare With OpenAI's?
OpenAI's OneGov offer is 10 points deeper and was on the schedule first. GSA announced it on September 10: 50% off token-based usage across ChatGPT models, including those in FedRAMP-authorized environments, with no platform-access fee, no minimum orders and no spend commitment, for a 27-month term starting October 1, open to federal, state, local and tribal governments. GSA's listing shows it running to December 31, 2028 (GSA OneGov).
Side by side, from the OneGov listing and each vendor's announcement:
| Term | Anthropic (Claude) | OpenAI (ChatGPT) |
|---|---|---|
| Discount | 40% off tokens | 50% off tokens |
| Commitment / minimum | $0 / $0 | $0 / $0 |
| Platform fee | Not stated on the listing | None, per GSA |
| Eligibility | Federal, state, local, tribal (Anthropic adds territorial) | Federal, state, local, tribal |
| Offer runs to | January 31, 2029 | December 31, 2028 |
| Old $1 offer ended | October 31, 2026 | September 30, 2026 |
Do not compare 40 to 50 and stop. A discount is a percentage of a list price, and the two vendors' list prices, token counts per task and model mix all differ. The comparison that matters is cost per completed task on your own workload, at both discounts. There is also a licensing wrinkle on OpenAI's side: Nextgov reported that OpenAI's blog cites a standard licensing fee of $15 per user per month, and Nextgov did not say whether the 50% applies to it. Ask OpenAI the same seat-fee question you ask Anthropic.
Both offers are open to states on MAS. Per GSA's listing, they are the only AI model offers on OneGov that extend to state and local entities; Google's Gemini for Government, Microsoft's Copilot offer and AWS's credits are federal-only (GSA OneGov). A state CIO who wants a discounted frontier model on a federal schedule now has exactly two choices on that page.
What Is Anthropic Claiming, and What Is Unverified?
Most of the adoption figures in the announcement are Anthropic's own. The company says about 500 US federal, state and local organizations use Claude, weekly Claude for Government active users are up about eightfold since February, and federal civilian usage grew about 20-fold in a year (Anthropic). It cites a Utah pilot in which 34 developers across 10 state agencies used Claude Code and 71% reported saving 20 or more hours over four weeks. These are self-reported pilot results from the vendor, published to sell a contract vehicle. Use them to justify a pilot of your own and measure the hours yourself.
The headline "$1 billion" in savings is likewise Anthropic's estimate of what the government saves, and a discount's savings depend entirely on how much gets bought. GSA says its OneGov program has saved more than $1.8 billion so far, about $1.6 billion from AI agreements, per The Next Web.
The political context has not gone away either. BusinessToday reports that a US appeals court upheld the Pentagon's "national security supply-chain risk" designation of Anthropic on September 25, that Anthropic sought a rehearing on October 13, and that the GSA agreement does not resolve that fight. The Next Web notes that HHS said it planned a sole-source contract for Claude models, and that Anthropic has told IPO investors US government actions could affect its business. Brief your counsel on the designation before you standardize on Claude. Background: Why the Government Cut Off Anthropic Overnight and Anthropic's S-1 Shows $16.3B H1 Revenue and Q2 Profit.
Why a Spend Cap Comes Before Anything Else
A metered contract without a ceiling is the most likely way this deal hurts you. Under the $1 price, usage had no budget consequence, so nobody tracked it. On November 1 every token has a price, and an agency's appropriation does not flex with a usage spike. We made the same argument when OpenAI's $1 offer ended, including the Antideficiency Act exposure of an open-ended meter: OpenAI's $1 Deal Ends Sept 30. Put a Ceiling on the Meter.
The controls are ordinary FinOps. Issue keys per program or office so spend can be attributed, route traffic through a gateway that enforces a monthly ceiling, and default to the cheapest model that passes your evaluation, with Opus reserved for the work that needs it. Caching repeated context cuts the input bill further, provided your prompts do not defeat the cache (Prompt Caching Savings: One Timestamp Turns 80% Into 6%). For tooling, see Best FinOps Tools for AI Spend: No Dashboard Fixes an Untagged Key.
The commercial Enterprise plan shows how quickly a seat-plus-usage bill can surprise an owner who budgeted for seats alone. Under the $1 price your AI line was close to zero, so most of the first metered invoice is money nobody budgeted.
What Should Agency and State CIOs Do Now?
This Week:
- Confirm with your contracting officer whether your Claude access lapses on October 31 or rolls to the consumption terms automatically, and on which vehicle. Get the answer in writing.
- Put a not-to-exceed amount on whatever order covers November, even a rough one. You can raise it once October's usage is priced.
- Ask Anthropic or your reseller, in writing, whether the 40% applies to the Claude Enterprise seat fee and whether the 20-seat minimum applies under OneGov.
This Month:
- Pull October's usage logs by office and model, and price them at the discounted rates in the table above. That is your baseline forecast.
- Run the same 20 to 50 real tasks through Claude and ChatGPT and record tokens per completed task. Multiply by each vendor's discounted price. Pick on that number.
- If you are a state, local or tribal buyer, check with your procurement office whether you can order from GSA's MAS under your state's cooperative purchasing rules before you plan a pilot around it.
Before Your Next Budget Submission:
- Move AI from a rounding-error line to a metered line item with a forecast range and an owner.
- Decide whether to standardize on one vendor or keep both behind a gateway. Both offers run into late 2028 or early 2029, so the decision you make now will hold for more than two budget cycles.
The Bottom Line
The $1 era was a land grab: GSA and the vendors priced frontier AI at nothing to get it onto desks. With OpenAI's offer in September and Anthropic's this week, both now charge for what you use, at a discount, for about 27 months. Cloud went through the same sequence a decade ago: free credits, then consumption bills, then FinOps teams hired after the first invoice nobody expected.
Set the not-to-exceed amount before November 1.
Continue Reading
- OpenAI's $1 Deal Ends Sept 30. Put a Ceiling on the Meter.
- Claude for Government Keeps the Only Copy of Each Chat on a Laptop
- Anthropic's S-1 Shows $16.3B H1 Revenue and Q2 Profit
- Best FinOps Tools for AI Spend: No Dashboard Fixes an Untagged Key
- LibreChat and Open WebUI offer ChatGPT Enterprise alternatives
- Claude Haiku 5.5 Charges 5x for Prompts Over 100K Tokens
