Curiosity · AI Model
OpenAI o1
OpenAI o1 is the September 2024 model that changed the LLM landscape. Rather than optimising for fast next-token prediction, o1 was trained with reinforcement learning to produce a long private chain of thought before answering — unlocking state-of-the-art scores on math, physics, and competitive coding while trading speed and cost for correctness.
Model specs
- Vendor
- OpenAI
- Family
- o-series
- Released
- 2024-09
- Context window
- 128,000 tokens
- Modalities
- text, code
- Input price
- $15/M tok
- Output price
- $60/M tok
- Pricing as of
- 2026-04-20
Strengths
- Landmark jump in math, science, and competitive-coding scores
- Far fewer 'silly' reasoning errors than GPT-4o on multistep logic
- Launched the template every major lab now follows for reasoning
- Strong refusal + safety behaviour from longer deliberation
Limitations
- Slow — responses can take 10-60 seconds even on moderate prompts
- Expensive per output token (includes hidden reasoning tokens)
- No tool use at launch; superseded by o3 / GPT-5 for agentic work
- No streaming visibility into its internal reasoning
Use cases
- Competition math and physics problem solving
- Research assistance for graduate-level scientific reasoning
- Complex code debugging and algorithm design
- Formal verification and proof outlining
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| AIME 2024 (math) | ≈83% | 2024-09 |
| Codeforces Elo | ≈1673 | 2024-09 |
| GPQA Diamond | ≈78% | 2024-09 |
Frequently asked questions
What is OpenAI o1?
OpenAI o1 is the first model in OpenAI's "o-series" of reasoning models, released in September 2024. It is trained with reinforcement learning to produce a long private chain of thought before answering, achieving large gains on math, science, and competitive coding.
How is o1 different from GPT-4o?
GPT-4o is a general-purpose multimodal model tuned for speed and breadth. o1 is a specialist reasoning model that deliberately thinks longer before answering — it is slower and more expensive, but much stronger on hard logical problems.
What tasks is o1 best for?
Competition math (AIME, IMO-style problems), advanced physics and chemistry, algorithmic coding challenges, formal proofs, and any domain where correctness outweighs latency.
Is o1 still worth using in 2026?
For most new projects, o3 or GPT-5 with reasoning enabled are better choices. o1 is historically significant and still available, but its successors are faster, cheaper, and support tool use.
Sources
- OpenAI — Introducing o1 — accessed 2026-04-20
- OpenAI — Pricing — accessed 2026-04-20