OpenAI's o1 was the first of a new kind of model - one that reasons before it answers. The specifics here are a September 2024 snapshot, but the decision it forces is evergreen: when is slow, careful reasoning worth it, and when do you want fast and cheap instead?
o1's "chain-of-thought" reasoning breaks complex problems into smaller steps, working through them like a human expert.
That makes it excellent for detailed, multi-step problems - and overkill for plenty of everyday tasks. Here's how I decide, after testing it on real work.
01The test: feedback on a book manuscript
I gave o1 two full chapters of my book and asked for line-by-line editorial feedback. It thought for 62 seconds before replying (unusually long - most responses came back in 5-10 seconds).
A glimpse of its visible reasoning: "Reading and analysing... mapping the request to OpenAI's policies... noting the engaging anecdotes and conversational tone... combining sentences for a more cohesive introduction... reducing redundancy and a grammatical issue... ensuring a clean, engaging storytelling flow." Dozens of steps like these, then a structured set of specific, located suggestions.
The feedback was incredibly detailed, and genuinely useful - I implemented several of the suggestions.
02Where o1 earns its slowness
The reasoning models shine when an expert could solve it - just not quickly.
Step-by-step challenges: debugging multi-step code, large sequential problems like physics or health-data analysis.
Working iteratively through options - product development, long-term strategy - testing approaches before committing.
It shows the reasoning, not just the answer - breaking a problem into steps and explaining each, which is where the learning is.
Where mistakes are costly - healthcare, aviation - the careful, step-by-step approach is worth the extra time.
When to stick with a faster model
For these, GPT-4o or Claude 3.5 Sonnet are the better call - speed beats depth.
03How to choose
Usage limits (Sept 2024 snapshot): OpenAI capped o1-preview at 30 messages a week (50 for o1-mini) on Plus/Team. Limits like these have eased over time, but reasoning models remain more expensive and rate-limited than fast models - plan accordingly.
Reasoning models are for problems an expert could solve - just not quickly.
For everything else - speed, creativity, conversation - a fast model wins. Match the model to the job.
Book a Focus Call