Claude Code found a $40 route to cheaper AI reasoning

Researchers from UMD, Google, Meta and other institutions did not hand-design another policy for AI reasoning. They let a coding agent search for one. According to The Decoder, the AutoTTS system used Claude Code to independently discover algorithms that decide how much compute a model should spend during inference.

That is a more exact problem than it may first sound. Modern systems increasingly do not ask a model once and stop. They can request multiple attempts, compare reasoning chains, vote across answers or terminate the process when extra samples no longer justify their cost. This layer is usually described as test-time scaling: performance is extracted not only from larger training runs, but from smarter model use at answer time.

Inside that space, AutoTTS found a method that, according to the supplied context, matches the accuracy of standard self-consistency while using about 70 percent less compute. Self-consistency is a known technique in which a model generates multiple reasoning chains and the most consistent answer is selected; the original method is described in Self-Consistency Improves Chain of Thought Reasoning in Language Models. Its weakness is not subtle: more attempts mean more tokens, more latency and a larger infrastructure bill.

That is why the most important detail may be the cost of the search, not just the compute reduction. The reported cycle took 160 minutes and cost $40. If that result holds across a broader set of tasks, this is not merely an optimization trick. It is a sketch of a cheap loop in which a coding agent writes, tests and selects algorithms that human researchers may not have proposed intuitively.

The institutional context matters, but it should not be treated as decoration. UMD, Google Research and Meta AI give the work weight, yet the technical message is broader: coding agents can search the design space of optimization policies, not just help implement them. In other words, the agent is not only an assistant accelerating a known procedure. In this case, it participates in choosing the procedure itself.

Precision still matters. The supplied context does not show that AutoTTS discovered a universal law of AI reasoning, or that the same algorithm will work equally well across all models and tasks. What it does show is concrete enough: in the reported experiment, it found a control algorithm that reduced compute compared with standard self-consistency while preserving accuracy. If such loops prove stable, an important part of future AI systems may be this quiet control layer: deciding when extra reasoning is worth paying for, and when it is just expensive repetition.