Google says teams of its AI agents checked a landmark math proof and rewrote 800,000 lines of code. It hasn't disclosed the cost
Google wants businesses and universities to pay for AI agents that work in teams on long, difficult jobs. On November 13 it published its biggest showcase yet for that pitch. Nearly all of it is graded by Google itself, and none of it comes with a price tag.
The post, from Google Research vice president Vahab Mirrokni, describes a feature called /teamwork inside Antigravity, Google's AI agent platform, which has paid tiers. It splits a big task across many Gemini 4 agents. Separate checker agents and outside tools then test each piece before it is accepted.
The headline claims are large. Google says the system produced a computer-checked version of the proof of the Weil conjectures, a landmark result in number theory and geometry. The proof was first completed by Pierre Deligne in 1974. Google says the new version runs to about 1.8 million lines and took 28,800 separate agent runs, "fully autonomously."
Google also says a two-week effort finished rewriting Zircon, the core of one of its operating systems, into Rust, a programming language designed to prevent common security bugs. An earlier effort had moved about a quarter of it; the two-week push added about 800,000 lines of Rust. Google says the result passed the kernel's own tests and booted on real hardware.
Why buyers should care
The math is the eye-catcher, but the code rewrites are the commercial story. Large companies and governments sit on millions of lines of aging software that is costly to maintain and hard to secure. Moving it to safer languages is slow, expensive specialist work.
If agent teams can do that in weeks, the market for software modernization changes. Google gave a second example: a rewrite of libgav1, a video decoder, into roughly 170,000 lines of Rust. Google says the first version ran four times slower than the original. After the agents were given speed-measurement tools, Google says it ran 20% to 30% faster, with identical output.
Google also reported gains on its own infrastructure. It says agent-suggested scheduling changes cut interruptions to some AI chip workloads by 24%. In chemistry, it says a 30-agent setup identified test molecules from lab scan data correctly 95% of the time on its first guess, nearly double a single agent.
Google is now using the results to recruit users. It is offering 10,000 academic researchers six months of free access to Antigravity's highest usage tier. Google has not said what that tier normally costs. Data Studios, a technology analysis site, estimated it at about $1,200 per researcher, or up to $12 million in total, based on Google's published price list. Google has not confirmed that figure.
What Google has not shown
The post gives no cost, computing time or chip count for the Weil project, or for most others. For a buyer, that is the key missing number. A rewrite that takes two weeks but burns an unknown amount of computing power cannot be compared with a human team's quote.
Outside checking is also thin so far. None of the early coverage, from Interesting Engineering, LavX News and Data Studios, cited an independent expert who had reviewed the results. Google's post links to research papers for several smaller results. It does not link to the Weil proof files or to the Rust code, so outsiders cannot yet rerun them.
That matters because "checked by computer" is a narrower promise than it sounds. The proofs were written for Lean, a tool that only accepts a proof if every logical step holds. But Lean checks the proof against a statement that a person or AI wrote down. If that statement is subtly wrong, the proof can pass while proving the wrong thing.
Google acknowledges this. Its post says the guarantee holds only once human mathematicians confirm that the formal statement matches the original problem. It does not say that step has been completed for the Weil work.
A day before Google's post, the Xena Project blog, run by mathematician Kevin Buzzard, set out a checklist for exactly this kind of claim. It does not mention Google. It warned that AI-written Lean code needs testing first, and that one must do work "beyond 'the Lean code compiles'." Its tests include confirming the statement matches the original and that the proof relies only on standard mathematical assumptions.
How to read it
Some of Google's claims are modest when read closely. The Weil result is a new computer-checked version of a 50-year-old proof, not new mathematics. A Navier-Stokes item in the post formalizes a recent published proof; it does not claim to solve the famous open fluid-flow problem. The Go-playing example shows a single game, which Google says was the only one played for the post. A survey of users of its paper-review tool had 70 respondents and is still running.
Others, if they hold up, are meaningful. Google says /teamwork helped prove a 2013 conjecture in information theory in full, and posted that proof as a research paper. Independent mathematicians can now test it.
For executives, the signal is about direction more than any single result. Google is betting that the next contest in AI is not which model answers a question best. It is which vendor can run many agents for days on one job and prove the output is right. Google has put out bold claims. The cost data and outside audits will decide whether they become a business case.