Steer GPT-5 Thinking Mid-Response to Save Time and Tokens
Intervene during the reasoning phase to steer the model's logic.
Watch the 'Thinking' phase preamble and provide mid-response instructions to correct the model's path before it generates the final answer.
The Scenario
You are asking the model to perform a deep web research task or a multi-step data analysis. You notice in the initial 'Thinking' preamble that it is about to use an incorrect source or method.
Before & after
Previously, you had to wait for the entire response to finish before prompting again to fix errors, often taking 10–15 minutes of back-and-forth.
Users can now give feedback or change priorities mid-thought, ensuring the final output is correct on the first try in about 2–4 minutes.
The Prompt
[DESCRIBE_COMPLEX_TASK_HERE] (Wait for the 'Thinking' plan to appear. If it looks wrong, type:) "Wait, adjust your plan: [INSERT_CORRECTION_HERE]. Ensure you prioritize [SPECIFIC_DETAIL] before proceeding."
GPT-5.4/5.5 Thinking now displays an upfront plan. If you see the model heading in the wrong direction during the 'thinking' phase, you can intervene immediately to save time and token usage.
Source
Model Release Notes | OpenAI Help Center"provide an upfront plan of its thinking, so you can adjust course mid-response while it’s working"