Thoughts Brewing Blog

AI Quick Tips 324: New OpenAI model - GPT 5.5

Written by Damien Griffin | May 9, 2026 4:59:59 PM

They aren’t slowing down the pace with these model releases. Not sure if you think that is a good thing or not, but by the time you decide, there will probably be another new model…

Anyway, OpenAI, the company that makes ChatGPT, has a new member of the GPT-5 model family - GPT-5.5, and according to their documentation, this model is a strong fit for -

  • Coding use cases
  • Tool-heavy agents
  • Grounded assistants
  • Long-context retrieval
  • Product-spec-to-plan workflows
  • Customer-facing workflows where execution quality and response polish are critical

 

New Model, New Prompts

They recommend looking at this model as a new model family and not a drop-in replacement for GPT-5.2 or GPT-5.4.

This mainly means that you should expect to either create new prompts for this model or, at least, re-write existing prompts so that they work well with the model.

 

GPT-5.5 works better with outcome-first prompts

Give it a clear goal and constraints, then let the model figure out the path forward.

WIth previous models (even GPT-5.4) you saw better results if you included step-by-step instructions. That’s generally not the case with GPT-5.5. This is part of why re-using your old prompts isn’t a good idea.

 

More Concise & Direct

The model is also more concise and direct by default. This can be good or bad.

It makes the model task-oriented and useful for production workflows but conversational experiences “may need explicit personality, warmth, rationale, and formatting guidance”

 

Prompt Recommendations

Here is some prompting guidance directly from OpenAI’s documentation:

  • State the expected outcome and success criteria.
  • Reduce or remove detailed step-by-step process guidance. Let GPT-5.5 choose the path unless the product requires that path.
  • Remove output schema definitions from the prompt where possible. Use Structured Outputs instead.
  • Optimize your prompt for caching: static parts first, dynamic parts last.
  • Drop the current date. The model is already aware of the current UTC date



There’s more to it (a lot more actually), but this is plenty to get you started.

 

Let us know what your questions are in the COMMENTS below ↓

 


 

Watch this video for a quick overview↓