X post claims NVIDIA, MIT cut coding-agent API costs by about half

Unconfirmed: an X post claims NVIDIA and MIT researchers cut coding-agent API costs by 50-54% by changing only the agent harness, and Automatica has not confirmed it.

@Mr_Buzzoni1 wrote on 2 November that "NVIDIA and MIT researchers pointed Karpathy's loop at the agent harness itself and cut the API bill by 50-54%" and that it was "same models, same benchmark. only the harness around Claude Code and Codex changed." The account had about 28,404 followers and the post had 92 likes when observed, but it remains a single non-institutional claim from X rather than confirmation from either institution.

In the same post, @Mr_Buzzoni1 described the loop as "propose -> implement -> review -> test -> keep or discard" and said "the 4 changes that survived" were:

• "edit a file and run its test in one tool call"
• "compact context only after a finished step, and only when it saves more than it costs"
• "archive huge outputs and pass on a short excerpt plus a handle"
• "hand log reading to a cheaper model, then check its evidence against the original"

The same post also says "the paper runs 15 pages," but does not link or name that paper, and the available evidence does not show comment from NVIDIA or MIT. The post includes no benchmark name, model versions, pricing basis, run count, or indication of whether the reported reduction is an average, median, or best case @Mr_Buzzoni on X1.

Those omissions matter because they leave the central performance claim impossible to reproduce from the public record. If the cost reduction came from fewer tool calls, lower context-token use, or routing some work to a cheaper model, the size of the gain would depend heavily on the benchmark mix, model pricing, prompt lengths, and stopping criteria. None of that is established by the post itself.

What would confirm the rumour is straightforward: a named paper or preprint with NVIDIA and MIT authors, plus the benchmark, model versions, and pricing basis used for the 50-54% figure, or a statement from NVIDIA or MIT acknowledging the work. What would seriously weaken or kill it is a denial from either institution, or a paper that does not contain the quoted result or the claimed setup. As of publication, neither company has publicly confirmed anything in the available evidence.