The Knob That Replaced Temperature
For years, prompt engineers talked about temperature the way photographers talk about aperture. It was the single dial that changed the character of a model's output, and learning to set it well was half the craft. That conversation is mostly over. On modern reasoning models, the setting that actually changes behavior is the effort level, and the useful question is no longer how creative should this be, but how much thinking does this task deserve.
OpenAI's ChatGPT now exposes a picker that reads like a budget menu: Instant, Medium, High, Extra High, and Pro, with the higher tiers reserved for paid plans and the top of the range running a frontier model at maximum reasoning. That structure is a confession about how these systems work. The model is not smarter in a different way at each setting. It is spending more tokens thinking, and you are paying for that in latency and money.
Why Effort Beats Temperature
Temperature controlled randomness, which mattered when models were mostly pattern completers. Effort controls reasoning depth, which matters now that models can work through a problem step by step before answering. On a hard multi-step task, turning temperature up does not help you. Giving the model room to think does.
This reframes prompt engineering from wording tricks to resource allocation. A vague prompt on a high effort setting wastes thinking on the wrong interpretation. A precise prompt on a low setting can fail simply because the model did not have room to work. The craft is matching the budget to the task, then writing an instruction that uses that budget well.
- Instant — lookups, rewrites, formatting, simple extraction. Speed is the whole point.
- Medium — everyday analysis, summaries, first drafts, moderate code changes.
- High — multi-step reasoning, tradeoff analysis, debugging, structured planning.
- Extra High — genuinely hard problems where a wrong answer is expensive.
- Pro — long-running, high-stakes work that can tolerate waiting.
How to Pick a Level Without Guessing
The lazy approach is to always pick the highest setting, and it is a mistake for two reasons. First, you pay for it in wall-clock time, which is the scarcest resource in an interactive workflow. Second, extra thinking on a simple task sometimes produces over-elaboration: the model invents constraints you did not ask for and buries the answer in caveats.
A better habit is to classify the task before you type. If a competent intern could answer from memory, use Instant. If it needs a checklist, use Medium. If it needs a checklist plus judgment about competing options, use High. Reserve the top tiers for decisions where the cost of being wrong dwarfs the cost of waiting.
Then validate the choice with a cheap test. Run the task at Medium and at High, compare the outputs, and look for a real difference. If High is not materially better, you have found free latency. This one comparison, done a few times on your own work, teaches you more than any general guide.
Writing Prompts for Low Effort
Low effort settings reward compression. The model has little room to deliberate, so the instruction should do the deliberating for it. State the exact output format, name the fields, give one example, and forbid everything else. A good Instant prompt looks like a form, not a conversation.
For example, instead of asking for a summary of a document, ask for five bullets, each under fifteen words, leading with the most decision-relevant fact. The model no longer has to guess what kind of summary you want, and it does not need to think hard to produce it. Constraints are not a limitation here; they are the mechanism.
One more low-effort trick: put the answer first. When the model has almost no room to think, telling it to lead with the result and then justify it keeps the output usable even if the reasoning is thin. You are optimizing for a correct first line, not a beautiful essay.
Writing Prompts for High Effort
High effort settings reward structure. When the model has room to reason, give it something worth reasoning about. Supply the criteria, the constraints, and the definition of a good answer, then explicitly invite it to compare alternatives before committing. Ask for the reasoning to be shown separately from the recommendation, so you can audit the logic without wading through it.
It also helps to ask for uncertainty. A high-effort prompt that ends with a request to flag assumptions and identify the weakest part of the argument produces far more useful output than one that simply asks for an answer. You are paying for deliberation, so spend it on self-critique rather than confidence.
The Cost Math Nobody Shows You
Effort levels are a pricing decision disguised as a settings menu. Higher reasoning burns more tokens, and on API workloads that multiplies quickly. The same task run at the top of the range can cost several times what it costs at the low end, and the quality difference is often invisible on easy work.
This is why the smartest teams route by effort the same way they route by model. Cheap classification first, expensive reasoning only where it changes the outcome. The two ideas are the same idea: stop paying frontier prices for tasks that never needed frontier thinking.
Prompt engineering used to be about finding the magic words. It is now about allocating a budget — and the teams that win are the ones that spend their reasoning tokens where the decision actually lives.
Putting It Together
Build a small map for your own work. List the tasks you repeat, tag each with the lowest effort level that reliably passes, and treat that as your default. Escalate deliberately when a task is high-stakes or when the cheap setting visibly fails. Review the map monthly, because the models improve and yesterday's High often becomes today's Medium.
The temperature era trained a generation of prompt engineers to obsess over wording. The effort era asks a more honest question: how much thought is this problem worth? Answer that first, and the wording gets easier.
