Reasoning effort becomes a control for LLM cost, time and reliability

Sebastian Raschka writes that almost 2 years have passed since OpenAI released o1, and that DeepSeek-R1 followed about 4 months later with an RLVR recipe. His new article explains how models learn low, medium and high reasoning effort modes, similar to current model families with multiple levels of compute effort.

Reasoning effort gives teams a knob instead of one default answer

Raschka frames a reasoning model as a model that produces an intermediate trace, works step by step and can correct itself along the way. Effort modes are not just longer answers. They are a choice about how much compute and time the model should spend before the final result.

In practice, this is a product interface over inference scaling. For trivial questions, high effort is waste. For code, math or costly decisions, low effort can be false economy.

Developers will tune thinking budget alongside latency

For application teams, this creates a new control layer. It is not enough to choose a model and write a prompt. Teams will have to decide which task classes get the cheap mode, which get higher effort and when a request should escalate.

That fits agent systems well. An agent can run cheaply through routine steps, then switch to higher effort while planning, repairing an error or before an action that changes data.

More thinking does not automatically mean more truth

Higher reasoning effort can improve performance on some tasks, but it does not guarantee correctness. A model can use a longer trace to defend the wrong direction more elegantly. Effort modes therefore need evals, latency measurement and cost control, not the comforting belief that longer thinking is quality by itself.

Good evals will decide when the model should slow down

The next work is mapping tasks to effort modes. Teams will need evals that show where higher effort actually improves accuracy and where it only burns tokens. Only then does reasoning effort become a reliable control, not decoration in an API.

Lilith's verdict

Reasoning effort is the gas pedal on an expensive engine. Keep it pressed all the time and you do not look smarter. You just leave a longer receipt for the same trip.