OpenAI · 2024 · 09 · 12 · Model · ~2 min read
OpenAI launched o1-preview
What's actually new
- It thinks out loud. The model spends seconds — sometimes a minute — working through a problem before it answers. You can watch the reasoning.
- Big jumps on hard tests. Maths-olympiad problems, coding contests, PhD-level science questions — much better than GPT-4.
- First AI where you can see the model think about its mistakes.
If you want more
Worth knowing
- 'Reasoning' is a friendly word for what's really a longer thinking trace. The model still makes things up. It still has a knowledge cut-off date.
- It's slow and expensive per question. Wrong tool for chat, summarising, or anything time-sensitive.
- Early access was thin. Many people only tried o1-mini, which is meaningfully weaker than the full o1.
Who should care
Developers and researchers working on hard problems where wrong answers cost real money. Students learning maths or coding. Anyone testing whether AI can help with serious science or engineering work.
What to do about it
Use thinking-style models for hard, well-defined problems where you're happy to wait 20-90 seconds for a better answer. Keep using fast models like GPT-4o for normal chat and everyday tasks.
Honest take
o1 was the moment AI shifted from a fast guesser that bluffed confidently to a slow thinker that admits hard problems are hard. That's a category change, not a feature update. The catch: it's the wrong tool for most everyday questions, and people who plugged it into existing workflows got a price-per-call shock. The bigger lesson o1 proved is that you can keep getting AI smarter by giving it more time to think — not just more data to train on.
Sources
Last verified · 2026 · 05 · 05 · Found a fact wrong? corrections@aguidetocloud.com