New Paper Proposes Formal Contract for Reasoning Effort in AI APIs

A new preprint on arXiv introduces a framework for formalizing the relationship between a user's request and the computational effort an AI model should expend in responding. The paper, which is still in early stages, argues that as AI models become more powerful and expensive to run, developers need a clearer, more standardized way to specify the desired level of "reasoning effort" for each API call.

The authors propose an "effort API contract" — a set of parameters and behaviors that would allow clients to signal whether they want a quick, low-cost answer or a deeper, more computationally intensive chain of thought. This would go beyond simple temperature or max-token settings, offering a more semantic control over how much "thinking" the model does.

For developers building on top of large language models, this could be a significant shift. Currently, managing the trade-off between response quality and cost often involves hacky workarounds, like adjusting prompts or iterating with different sampling parameters. A standardized contract would make this more predictable and portable across different model providers.

The paper stops short of proposing a specific implementation, instead focusing on the theoretical underpinnings and potential design space. It's an open question whether major AI labs will adopt such a standard, but the research reflects a growing concern: as models get more capable, the cost of their "thinking" becomes a first-class concern for developers.

For now, this is a research proposal, not a shipping product. But it's a glimpse into a future where AI APIs might expose a dial for reasoning effort — and where developers can make more intelligent decisions about how much they're willing to pay for the right answer.