Article 341: The US Government Just Moved From $1-a-Seat Claude to a 40% Token Discount. That Is the Whole Pricing Debate in One Contract
The General Services Administration expanded its OneGov agreement with Anthropic this week. Federal agencies - and according to the announcement, state, local and tribal governments too - now get 40% off token-based usage across Claude models, on consumption-based terms, with no minimum purchase.
It replaces an offer that was already on life support: $1 per user for Claude for Government, extended once and due to expire on 31 October. Google's competing Gemini deal, at $0.47 per agency, runs to 15 November with no announced successor.
Why this is bigger than a discount
The interesting thing is not the 40%. It is that the federal government has now adopted the pricing model its own analysts have spent a year criticising.
Gartner has said consumption-based pricing makes AI costs hard to predict, because vendors do not give customers transparent visibility into how token consumption is metered and billed. Forrester's Tracy Woo put the operational version of that in August: "A small group of power users can drive a disproportionate share of costs, turning successful AI adoption into both a productivity win and a budgeting challenge."
Both are correct, and both descriptions apply to a government agency exactly as they apply to a company. The difference is that a department that overspends a token budget in March has to explain it to appropriators, not a CFO.
The tell is in the sequence
Look at what GSA did in order. It extended the promotional deal once. It then replaced the flat promotional price with a discounted metered price, matching OpenAI, which made the same switch last month. That is not a procurement accident. It is what happens when the pilot phase ends and someone asks how much this will cost at scale: the honest answer is per-token, because that is how the vendors charge and how their own costs work.
The 40% is the compensation for taking on that variance. The number is generous. It is also a discount off list price, not a ceiling, and there is no minimum purchase, which cuts both ways - no commitment, and no floor on the bill either.
What agencies now have to build
If you are running Claude at an agency, the job that just appeared is not "adopt AI". It is metering. Per-group budgets, spend visibility at the team level, and somebody whose job is to notice when a session's cost is ten times the median. Which, awkwardly, is roughly what Anthropic also shipped this week for enterprises: pooled group budgets and reports that attach list-price spend to workstreams.
The vendors are selling both the consumption and the odometer. That is not a conspiracy, it is just what the market looks like when nobody can price a token in advance.
Key Takeaway: Metered pricing won because it is the only model that survives contact with usage that is genuinely unpredictable. The promotional $1 seat did its job - it got tools into agencies for a year - and it had to end, because a per-seat price on an agent that runs for hours is a subsidy. The real question for every public-sector buyer is not whether 40% is a good deal. It is who in the building is accountable for the meter, and whether they can see it before the invoice arrives rather than after.