Prompting reasoning models
The single biggest change in prompting practice is that reasoning control moved out of your prompt text and into an API parameter. Most guides have not caught up.
Stop telling them to think step by step
This is the clearest case of advice that expired. All three major vendors now say, in their own documentation, that explicitly instructing a reasoning model to reason is redundant at best.
OpenAI puts it plainly: since these models perform reasoning internally, prompting them to think step by step or explain their reasoning is unnecessary. Google tells developers migrating to Gemini 3 to drop the chain-of-thought scaffolding they previously used to force reasoning, and to raise the thinking level instead. Anthropic now frames manual chain-of-thought as a fallback for when thinking is switched off, and advises preferring general instructions over prescriptive steps, because the model's own reasoning frequently exceeds what a human would prescribe.
Cross-vendor parameter table
Verified against vendor documentation on 27 July 2026. These change often, so check the source before relying on it.
Two details worth knowing. Anthropic's older extended-thinking configuration is not merely discouraged on current models, it is rejected: requests using it return an error rather than falling back. And OpenAI separates how long the model thinks from how long its answer is, so you can ask for deep reasoning and a short response at the same time. That pairing is underused.
Google also advises removing low temperature settings when moving to Gemini 3 and using the default instead, which reverses the old habit of setting temperature near zero for consistency.
Why they feel slow, and when to accept it
Reasoning tokens are billed as output on every major platform, and they are generated before you see anything. That is the whole explanation for both the latency and the bill. A high effort setting on a simple extraction task is pure waste in two currencies at once.
OpenAI's own guidance is to treat the effort setting as a tuning knob rather than the primary way to recover quality. If the answer is wrong because the brief was ambiguous, more thinking produces a more elaborate wrong answer. Fix the brief first, then tune effort.
Start one level below what you assume
When migrating from an older model, OpenAI suggests keeping your current effort as a baseline and comparing one level lower. The cheaper setting often matches.
Reserve maximum effort for genuinely hard work
Multi-step analysis, tricky debugging, work where being wrong is expensive. Not summarisation, classification or reformatting.
Re-tune after every model upgrade
Anthropic states this outright: run a fresh effort sweep against your own evaluations rather than carrying settings across. A setting that was right in spring may be wasteful now.
Separate thinking length from answer length
Where the platform supports it, ask for deep reasoning and a concise response rather than accepting a long answer as the price of good thinking.
Stop shouting at them
Older models needed pushing to use a tool or follow a rule, so people learned to write things like “CRITICAL: you MUST always call this tool.” Current models have the opposite failure mode. Anthropic reports that its recent models overtrigger, and advises dialling that language back to ordinary phrasing.
The same shift applies to repetition. OpenAI's current guidance is to state each instruction once, and it reports that removing repeated instructions and simplifying tool descriptions improved both task performance and token efficiency in its own testing. Treat those specific figures as vendor-reported rather than independently confirmed, but the direction is consistent across all three vendors.
Write this
- State each rule once, clearly
- Give the task, the constraints and the output format
- Use the effort parameter for depth
- Say what “done” looks like
Not this
- ALL-CAPS emphasis and “you MUST”
- The same instruction restated three ways
- “Think step by step” on a reasoning model
- Temperature copied from an older model's recipe
Sources
- OpenAI: reasoning best practices — explicitly advises against chain-of-thought prompts
- OpenAI: reasoning guide — effort as a tuning knob, not a quality fix
- Anthropic: effort parameter — the effort scale and what it controls
- Anthropic: prompting best practices — overtriggering, manual chain-of-thought as fallback, re-tuning after upgrades
- Google: Gemini 3 developer guide — thinking_level, and removing low temperature settings