Utah ran a Claude Code pilot with 34 state developers

Boris Cherny, who works on Claude Code at Anthropic, posted the results of a four-week Claude Code pilot run by the state of Utah. The pilot covered 34 developers across 10 state agencies. According to Cherny, 80% had a working environment within 30 minutes, and 71% reported saving 20 or more hours by the end of the four weeks.

Boris Cherny

@bcherny

X

80% had a working environment within 30 minutes, 71% reported saving 20+ hours by the end of the four-week pilot.

That is the whole data set as posted. There is no breakdown by agency, no description of what the developers were working on, and no detail on how the hours saved were measured or collected.

Read the two numbers differently

The hours number and the setup number are not the same kind of claim. Self-reported time savings from a pilot run with a vendor are the softest evidence there is. Developers who volunteered for a tool trial, were asked at the end whether it helped, and knew the answer would be read, are not a controlled sample. Twenty hours over four weeks is roughly an hour a day, which is the sort of estimate people round to when they feel faster.

The 30 minute setup figure is harder to fudge. Either the install worked on a state-issued machine behind state network policy or it did not, and 80% of 34 developers getting to a working environment inside half an hour is a specific, falsifiable thing. Anyone who has tried to get a new CLI tool approved and running inside a government IT environment knows that is the part that usually kills a pilot before anyone writes a line of code.

The context around the post

The numbers landed alongside Anthropic's announcement that it is expanding its partnership with the GSA. Under the new OneGov agreement, the company says it is bringing Claude to federal, state, local, tribal and territorial governments at 40% off.

So this is a vendor publishing a favourable pilot result the same week it announces discounted government pricing. That does not make the numbers wrong. It does mean the useful signal here is procurement, not productivity. The interesting question for working engineers is not whether 34 people in Utah felt faster, it is whether agent tooling can clear government environment and security review at all. On this evidence, for one state, it cleared setup in 30 minutes for most of the people who tried.