Activereasoningllm Proprietary

o1

by OpenAI· Released September 2024· Cutoff October 2023

o1 is a large language model by OpenAI designed for complex reasoning tasks, using chain-of-thought reasoning to solve difficult problems in science, coding, and math. It is the first model in the o-series, emphasizing deep thinking and accuracy over speed. o1 is available in preview and mini variants, with o1-preview being the full reasoning model and o1-mini a faster, cheaper option for coding and STEM.

Official Site API Docs

Input cost

$15.00 per 1M tokens

Output cost

$60.00 per 1M tokens

Context window

128K tokens

Max output

32K tokens

Modalities

text

License

proprietary

Capabilities

Chain-of-Thought ReasoningCode GenerationMath Problem SolvingSTEM KnowledgeFunction CallingJSON ModeStreaming

Best For

Complex reasoning tasks requiring deep analysis, such as advanced math, science, and coding challenges.

Strengths

State-of-the-art reasoning capabilities, outperforming GPT-4o on many benchmarks
Strong performance on coding and math tasks (e.g., 83% on AIME math, 89th percentile in Codeforces)
Ability to think step-by-step and self-correct
Large context window (128K tokens) for handling long documents

Limitations

Slower response times due to internal reasoning process
Higher cost compared to GPT-4o and GPT-4o-mini
No vision or multimodal capabilities (text-only)
Limited availability (preview stage) and may have usage rate limits

Use Cases

Solving complex math problems

Advanced coding and algorithm design

Scientific research and hypothesis generation

Legal or financial analysis requiring deep reasoning

Debugging and optimizing code

Educational tutoring for STEM subjects

Data analysis and interpretation

Improvements Over Previous Model

First model in the o-series, so no direct predecessor; compared to GPT-4o, it introduces chain-of-thought reasoning for complex tasks
Significantly better performance on math (AIME: 83% vs GPT-4o's ~12%) and coding (Codeforces rating 1807 vs GPT-4o's 808)
New reasoning paradigm that allows the model to think before answering, improving accuracy on multi-step problems
Higher output token limit (32K vs GPT-4o's 16K) for longer reasoning chains

Back to all models

Improvements Over Previous Model

First model in the o-series, so no direct predecessor; compared to GPT-4o, it introduces chain-of-thought reasoning for complex tasks

Significantly better performance on math (AIME: 83% vs GPT-4o's ~12%) and coding (Codeforces rating 1807 vs GPT-4o's 808)

New reasoning paradigm that allows the model to think before answering, improving accuracy on multi-step problems

Higher output token limit (32K vs GPT-4o's 16K) for longer reasoning chains