On October 27, 2026, the US General Services Administration announced that its OneGov agreement with Anthropic has been rewritten to a token-metered contract: a 40% discount on consumption pricing for every Claude model available to government, no minimum spend, open to federal, state, local, tribal and territorial agencies, inside the FedRAMP High environments authorised for sensitive-unclassified workloads. The old deal — Claude for Government and Claude Enterprise at one dollar per user per year — expires on Saturday, October 31.

That dollar-per-user line, which generated a fortnight of headlines when it landed in August 2025, got pushed from a September 30 cutoff to October 31 after Anthropic extended it by a month in mid-September. It was the second extension of a pilot that was never going to be the commercial terms. OpenAI switched its own OneGov agreement to usage pricing in September. Google Gemini is still at $0.47 per agency through November 15 — a number carefully chosen to be less than one dollar.

What the deal actually is

The 40% is off list, applied to the per-token price for Claude Enterprise, Claude for Government and Claude API. There is no annual commitment, no floor, no per-seat metric. Agencies pay for the tokens they burn, discounted, and they can buy directly from Anthropic, from an authorised reseller, or through the three major cloud marketplaces — "in the coming months", per Anthropic's post, so that path isn't live yet. The discount applies inside Anthropic's FedRAMP-High environments. State, local, tribal and territorial governments are included on the same terms, a scope widening that neither the 2025 $1 pilot nor OpenAI's September successor carried.

The duration isn't in the GSA release. FedScoop's reporting, which the Loop couldn't match against the GSA primary text, puts the window at through December 31, 2028. If accurate, that gives agencies a 26-month purchasing path at the new terms — long enough to run a budget cycle through, short enough to leave room for a 2029 repricing.

The scope widening is the consequential part. OneGov deals until now have been federal-focused with "approved for state and local via GSA MAS" as a secondary note. The Anthropic announcement puts states and localities on the same discounted sheet from day one. A 50-state buying posture matters more than a 40% number when the number is applied to the same per-token prices commercial customers also negotiate off.

The savings math is doing a lot of work

Both sides came out with their own big-dollar figure. Anthropic's post said the new terms "could save the government up to $1 billion." GSA's rolled the whole programme up: "more than $1.8 billion in total savings from OneGov" since launch, with "approximately $1.6 billion from super intelligence agreements." Neither number carries a baseline. Savings against what list — the retail API price, the previous commercial contract, the hypothetical consumption at full price if agencies use Claude at the volume Anthropic hopes they will?

That is a sentence that costs a billion dollars. The GSA figure also folds in whatever OpenAI and Google agreed under the same programme, so Anthropic's $1 billion contribution is a subset of the $1.6 billion attributed to "super intelligence." The arithmetic hasn't been shown; the aggregation hasn't been audited; the comparator hasn't been named.

Neither release names an official. There's no quote from Ed Forst, the GSA Administrator the Senate confirmed last October on the AI-adoption platform he campaigned on. There's no quote from Michael Rigas, who was Acting Administrator when the August 2025 $1 deal landed and told reporters at the time the US was "setting the standard for how governments adopt AI." There's no quote from Josh Gruenbaum, the Federal Acquisition Service commissioner who called the original deal "another major step in the AI-driven transformation of government." The press release reads like what it is — a terms sheet, not a pitch.

What the numbers Anthropic did share actually show

The adoption claims in Anthropic's post are specific enough to be checkable at the margins:

• About 500 US government organisations — federal, state, local — now use Claude.
• Weekly users of Claude for Government are up 8× since February.
• Civilian-federal usage is up about 20× year-on-year.
• A four-week pilot with the State of Utah involved 34 developers across 10 state agencies using Claude Code; 71% reported saving more than 20 hours over the period.

The Utah pilot is the only one that comes with a methodology — small enough to be plausible, specific enough to be audited, framed as a developer-productivity claim rather than a vague "transformation." The 8× and 20× figures are off bases Anthropic hasn't disclosed, which is the sort of multiplier you cite when the absolute numbers are small.

The analyst view on consumption pricing in government

Trade press coverage flagged the gap between a 40%-off headline and what agencies can actually budget against. Tracy Woo, principal analyst at Forrester, told FedScoop earlier this year that "a small group of power users can drive a disproportionate share of costs" under consumption terms, and Gartner's analysts have described per-token pricing as "difficult to predict costs" against because vendors aren't transparent about how prompts and context windows multiply token counts in production use.

That is the second-order problem the switch creates. The $1 pilot was financially unforecastable in the simplest possible way — nobody knew what their eventual costs would be when the pilot ended, because the pilot wasn't a cost. The 40%-off consumption deal is unforecastable in a more respectable way: agencies can measure costs, but the measurement includes variables (prompt design, retrieval context, agentic retries) that even Claude's own account teams don't model consistently. A government CFO building an FY27 line item against this contract is going to want more than a token-rate card.

How we got here, in dates

• August 12, 2025: GSA announces the first OneGov deal with Anthropic. $1 per Claude for Enterprise / Claude for Government user, approved for the executive branch, pending approvals for Congress and the judiciary. Rigas and Gruenbaum give the sunny quotes.
• September 17, 2026: Anthropic extends the $1 pilot through October 31, buying another month after the original September 30 cutoff. Neither GSA nor Anthropic answers reporters' calls.
• September 2026: OpenAI switches its own OneGov agreement from $1-per-agency to a usage-based discount, about a month before Anthropic.
• October 27, 2026: GSA and Anthropic jointly announce the consumption-based successor. 40% off tokens, no minimums, SLTT included.
• October 31, 2026: the $1 pilot formally ends. Any agency still on the old terms moves to the new ones or pays retail.
• November 15, 2026: Google's current OneGov terms for Gemini ($0.47 per agency) expire. Watch for a parallel move to usage pricing.

What to watch

1. Whether the FedScoop through December 31, 2028 duration is in the signed terms. The GSA release doesn't name a date; Anthropic's doesn't either. If 2028 is correct, agencies have two full budget cycles to spend into the discount. If the term is shorter and the discount steps down after year one — a common consumption-contract shape — the headline number is softer than it reads.
2. Whether other OneGov vendors follow. OpenAI already did; Anthropic just did; Google's deal expires in three weeks. If Microsoft's Copilot pricing stays flat-per-seat while the model-API vendors move to tokens, agencies will have to decide at FY27 budget time whether they're procuring intelligence or seats.
3. Whether GSA publishes the baseline for the $1.6 billion "super intelligence" savings figure. Right now it's a press-release aggregate with no definition and no comparator. A procurement-economics-of-AI story is only as sturdy as its baseline, and this one doesn't have one yet.
4. Whether any single agency puts a concrete FY27 line item against the Claude contract. The Utah pilot was 34 developers across 10 agencies for four weeks. A single federal CFO publishing a numbered FY27 ask — "Claude token spend: $X million" — would turn this deal from a terms sheet into a precedent.

Through all of this, the structural shift is small but real: the Anthropic–GSA relationship stopped being a political photo op and started being a line item. The $1 era is the one that ends on Saturday. The arithmetic is now visible, if you can get an agency to show it to you.