o1
OpenAI's first-generation deep-reasoning model that thinks before answering
Context window
200,000 tokens (100,000 max output)
Input / 1M tokens
$15.00
Output / 1M tokens
$60.00
Provider
OpenAI
Cached input $7.50 per 1M. o1 is described in OpenAI docs as the 'previous full o-series reasoning model'; it remains available but has been superseded by newer reasoning models. Output tokens include billed internal reasoning tokens. · Data verified 2026-07-02
o1 is OpenAI's original o-series reasoning model, released in full form on December 17, 2024 (the o1-preview appeared September 2024). It produces a long internal chain of thought before responding, making it strong on hard math, science, and competitive programming. It has a 200,000-token context window with up to 100,000 output tokens (including billed reasoning tokens), accepts text and image input, and has a knowledge cutoff of October 1, 2023. Benchmark highlights include 74% on AIME 2024 (single sample), 77.3% on GPQA Diamond (exceeding PhD-level humans at ~69.7%), and an ~89th-90th percentile Codeforces ranking. It is priced at $15 input / $60 output per 1M and is now a previous-generation reasoning model.
Capability index
Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.
How to access it
Available in the OpenAI API via model id 'o1' (snapshot 'o1-2024-12-17'). The 'o1-preview-2024-09-12' snapshot is deprecated. Positioned as the previous full o-series reasoning model; still usable but superseded by newer reasoning models.
Strengths
- ✓Deep step-by-step reasoning for hard math and science problems
- ✓Strong competitive-programming performance (~89th-90th percentile Codeforces)
- ✓Exceeds PhD-level human accuracy on GPQA Diamond (77.3% vs ~69.7%)
- ✓Large 200K context and 100K output budget for long reasoning chains
- ✓Multimodal text + image input
Best for developers who...
When to choose it (and when not to)
Reach for o1 when...
- →Hard, multi-step math, science, and logic problems that reward deliberate reasoning
- →Competitive programming and algorithmic problem solving
- →Existing o1-based pipelines already validated for reasoning tasks
Look elsewhere if...
- ✕Latency- or cost-sensitive tasks (it is slow and expensive: $15/$60 per 1M)
- ✕Simple or purely conversational tasks (use GPT-4o or a newer general model)
- ✕New projects that could use newer, cheaper, stronger o-series/GPT-5 reasoning models
How to use it
- ›Keep prompts simple and direct - avoid explicit 'think step by step' instructions; o1 reasons internally by design
- ›Do not add few-shot chain-of-thought examples; they can hurt reasoning-model performance
- ›Budget generously for output/reasoning tokens on hard problems
- ›Provide only the essential context; state the goal and constraints clearly
Quickstart
Pythonfrom openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="o1",
messages=[{"role": "user", "content": "Prove that the square root of 2 is irrational."}],
)
print(response.choices[0].message.content)Pin snapshot 'o1-2024-12-17'. o1 reasons internally, so avoid chain-of-thought prompting; expect higher latency and token usage.
API model id: o1
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| AIME 2024 | 74% | Single sample per problem (pass@1); rises to ~83% with 64-sample consensus, per OpenAI. |
| GPQA Diamond | 77.3% | Zero-shot pass@1, exceeding PhD-level human experts (~69.7%), per OpenAI. |
| Codeforces | ~89th percentile | Competitive programming; OpenAI reports Elo 1807, better than ~93% of competitors. |
Compare o1 with
o1 vs GPT-5
OpenAI - 400,000 tokens (128,000 max output) ctx
o1 vs GPT-4o
OpenAI - 128,000 tokens (16,384 max output) ctx
o1 vs Claude Opus 4.7
Anthropic - 1M ctx
o1 vs Gemini 2.5 Pro
Google DeepMind - 1,048,576 tokens (1M) input; up to 65K output ctx