o1
o1 is a large language model developed by OpenAI, designed for complex reasoning tasks. Its defining characteristic is an extended internal chain-of-thought reasoning process performed before generating a response, enabling strong problem-solving performance in domains requiring multi-step derivation such as mathematics, science, programming, and logical analysis. o1 belongs to OpenAI's reasoning model series and is positioned distinctly from the GPT series, which is oriented toward general-purpose conversation.
Core Facts
Reasoning models, represented by OpenAI o1, DeepSeek-R1, and QwQ, have chain-of-thought expansion as their core mechanism, generating a large number of intermediate reasoning steps before producing a final answer.
O3 belongs to OpenAI's o-series reasoning models and introduces a Chain-of-Thought mechanism for multi-step internal reasoning.
OpenAI's o1 series models internalize Chain of Thought as a core mechanism of model training, using reinforcement learning to teach the model to automatically unfold detailed reasoning processes.
Timeline (last 90 days)
OpenAI's o1 model introduced reasoning capabilities, and AI began to break down problems, make plans, and self-evaluate, at the cost of longer wait times and more token consumption.
Exclusion criteria for self-hosting being unsuitable: OpenAI monthly spending below $200, a need for o1-level complex reasoning, no GPU and LLM operations experience, or inability to accept round-the-clock incident response.
The third wave of retirement occurs on November 19, 2026, when o1, o1-pro, and another five o-series models will be retired together.
Microsoft Azure OpenAI and Foundry services will retire multiple models including o1, o1-pro, codex-mini, and Claude Haiku 4.5 in November 2026.
The routing layer can be divided into three tiers: the local/edge tier (such as Llama 3 8B, Mistral 7B, cost about $0.05–$0.15/1M tokens, latency <250ms), the mid-tier hosted tier (such as Claude 3.5 Haiku, GPT-4o-mini, cost about $0.25–$0.60/1M tokens, latency <800ms), and the frontier reasoning tier (such as Claude 3.5 Sonnet, o1, GPT-4o, cost about $3.00–$15.00/1M tokens, latency 2s–8s).
Under the same hardware conditions, the fast model qwen2.5-coder:14b has a latency of about 2 seconds, about 500 tokens per answer, and a cost of about $5 per million answers; the thinking model deepseek-r1:14b has a latency of about 15 seconds, about 3000 tokens per answer, and a cost of about $30 per million answers; o1-class reasoning models have a latency of about 45 seconds, about 15000 tokens per answer, and a cost of about $150 per million answers.
Task routing dividing line: code completion, simple edits and formatting go to fast models; debugging, refactoring and multi-file changes go to thinking models; architecture decisions, mathematical proofs and new algorithms use o1-level reasoning models; summarization, extraction and classification use fast models.
Models with reasoning as a core training objective, such as o1/o3, have also shown notable capability leaps in non-mathematical domains such as medical diagnosis and legal analysis.
The cost per million tokens for OpenAI's high-end reasoning models (such as the o1 and o3 series) can be 10 to 50 times that of lightweight models (such as GPT-4o mini).
OpenAI's letter-suffixed versions such as o1 and o3 sometimes represent an entirely new reasoning paradigm rather than a simple iteration.
32 more timeline events
All Facts (20)
Reasoning models, represented by OpenAI o1, DeepSeek-R1, and QwQ, have chain-of-thought expansion as their core mechanism, generating a large number of intermediate reasoning steps before producing a final answer.
90%VerifiedO3 belongs to OpenAI's o-series reasoning models and introduces a Chain-of-Thought mechanism for multi-step internal reasoning.
90%VerifiedOpenAI's o1 series models internalize Chain of Thought as a core mechanism of model training, using reinforcement learning to teach the model to automatically unfold detailed reasoning processes.
90%VerifiedOpenAI's o1 series represents a new paradigm called "test-time compute"
80%VerifiedThe OpenAI o1 series models were released in September 2024, representing a paradigm shift in LLMs from fast, intuitive responses to deep reasoning.
80%VerifiedOpenAI的o1系列模型和DeepSeek-R1采用了类似Chain-of-Thought的显式推理机制
80%VerifiedIn early 2025, DeepSeek-R1 achieved reasoning capabilities comparable to OpenAI o1 at a training cost of approximately $6 million.
80%VerifiedGoogle's Gemini, Anthropic's Claude, and OpenAI's o-series models have all implemented different forms of extended thinking.
75%VerifiedThe reasoning paradigm is essentially the application of reinforcement learning to the post-training stage of language models, allowing the model to autonomously explore reasoning paths and receive reward feedback based on answer correctness, first publicly demonstrated in the o1 series of models.
75%VerifiedOpenAI has introduced a 'Thinking Budget' mechanism in its o-series reasoning models, which can dynamically control the length of the internal reasoning chain.
75%VerifiedThe number of internal reasoning tokens in OpenAI's o1/o3 series models is often several times to dozens of times that of the final output.
65%VerifiedOpenAI's naming system shifted during the GPT-4 era from decimal version numbers to letter suffixes (such as 4-turbo and 4o) to identify architectural optimizations.
65%VerifiedOpenAI's o1 model was released in September 2024; through reinforcement learning it was trained to perform internal reasoning before outputting a final answer, with intermediate reasoning steps existing as special tokens that are invisible to users.
65%VerifiedChain-of-Thought（思维链）提示技术通过强制模型先输出中间步骤来部分缓解自回归模型在逻辑推理中的缺陷
65%VerifiedIn the o1 series models, OpenAI internalized the chain-of-thought mechanism as a spontaneous model behavior and introduced the concept of Reasoning Tokens.
65%VerifiedOpenAI's o-series (o1, o3) are specifically reinforcement-trained for chain-of-thought reasoning capabilities, incorporating Monte Carlo Tree Search (MCTS) and process reward models.
65%VerifiedThe o1 model performs internal chain-of-thought reasoning before answering, far surpassing GPT-4o on tasks such as mathematics, programming, and scientific reasoning.
65%VerifiedGoogle's Gemini 2.0 Flash Thinking and models such as DeepSeek-R1 have followed the test-time compute scaling technical path of o1.
65%VerifiedReasoning models (such as OpenAI's o1/o3 series, Anthropic's Claude with extended thinking, and DeepSeek-R1) output their thinking process (thinking tokens) before generating the final answer.
65%Unverifiedo1系列于2024年9月发布，引入了思维链推理机制
90%