OpenAI gives 23 million US public servants ChatGPT at $0
Two OpenAI announcements landed within minutes of each other this afternoon: a government-wide procurement deal that waives the license fee entirely, and a data-analysis agent aimed at every company that already pays someone to build dashboards.
OpenAI and the General Services Administration signed a 27-month OneGov agreement that drops ChatGPT's $15-per-user monthly license fee to $0 and cuts usage costs in half for federal, state, local, and tribal governments. The deal runs from October 1, 2026 through December 31, 2028, and OpenAI says it expands eligibility from the roughly one million government employees covered by existing agreements to a US public-sector workforce of about 23 million people. Every verified government entity also gets approved access to Daybreak Blue, OpenAI's cyber-defense tier, at 50% off commercial pricing, and can request Daybreak Red for vulnerability research and red teaming at list price. GSA puts OneGov's total savings near $1.7 billion, $1.4 billion of it from AI tools used by about 3.5 million federal employees.
The interesting part is the timing. Last year's federal offer was a novelty; this one arrives as those deals expire and agencies have to decide whether to keep paying. A $0 license with no minimum commitment is the classic land-and-expand structure — the real revenue is metered usage, and the lock-in is that 23 million people get used to asking a model first. It also quietly makes OpenAI the default answer for the state and local governments that were never going to run a procurement bake-off, and the cyber-defense bundling gives the whole thing a security rationale that is much easier to defend at a city council meeting than "productivity."
OpenAI also shipped a Data agent inside ChatGPT Work that connects to Snowflake, Databricks, BigQuery, Redshift, ClickHouse, MongoDB and Datadog, then turns plain-language questions into interactive dashboards. It pulls business definitions from semantic layers like dbt, Snowflake Horizon and Databricks Genie, enforces the connected account's existing row- and column-level permissions, and can write results back into Tableau, Power BI, Sigma or ThoughtSpot. OpenAI says nearly all of its product team and over two-thirds of its go-to-market staff already use the internal version, which runs across 600 petabytes and 70,000 datasets.
What OpenAI did not ship is a benchmark. VentureBeat pressed the company on accuracy and got process instead: an internal comparison against OpenAI's own data tools, no published score — this in the same week Databricks published results claiming its retriever matches Claude Sonnet 5 and GPT-5.6 Luna quality at more than twice the speed. For a category where correctness is the entire product, "we hill-climbed it internally" is a weak answer, and enterprise buyers should treat the missing number as the story.
Magic, the coding-model lab, says its pretraining recipe is now more than 10x more compute-efficient than leading open-weight base models. It claims to match DeepSeek V4 Pro Base using roughly 50x fewer FLOPs — about half of GPT-3's pretraining compute, or around $500,000 on GB200 hardware — and that a 10x larger run at about $4 million beat every publicly available open base model on perplexity evals. The gains came from dozens of multiplicative changes across architecture, optimizer, training objective and data curation, each validated by training three models spanning two orders of magnitude of compute. The catch: these are self-reported numbers on a blog, with no released checkpoint yet, so the claims stay unproven until the model ships.
What to watch: whether other vendors respond to the OneGov terms before the October 1 start date — GSA has explicitly invited them to.
If your agency or employer gets ChatGPT for free, does that change how much you trust its answers? Tell us in the comments.
