OpenAI o1 ​​和 o3 推理模型解释

OpenAI o1 and o3 are the first two generations of what OpenAI called its o-series, or reasoning models. Like earlier GPT models they are large language models: neural networks trained to predict the next token, a word or word fragment, of text. The difference is in what happens before the answer. A reasoning model first generates a long internal chain of thought, a running scratchpad of intermediate steps, checks and false starts, and only then writes its visible reply. OpenAI's announcement of o1 summarized the idea as a model that thinks before it answers. The thinking is produced as reasoning tokens, which OpenAI's API documentation says are not shown to developers but still take up room in the context window and are billed as output tokens.

The chain of thought is not just prompted; it is trained. OpenAI says o1 was trained with a large-scale reinforcement learning (RL) algorithm, a method in which a model tries many attempts and is rewarded for good outcomes rather than shown the correct text to copy. Through RL, according to the September 2024 announcement, the model learned to break hard steps into simpler ones, notice and correct its own mistakes, and switch strategies when an approach was failing. The company reported that o1's performance improved smoothly with more RL training compute and with more test-time compute, meaning more time spent thinking per question. That second axis was the headline: accuracy could be bought at inference time as well as by building a bigger model.

The timeline is short. OpenAI released o1-preview and o1-mini on 12 September 2024, followed by the full o1 in December 2024 alongside a system card describing its safety testing. On 20 December 2024 it previewed o3, and the ARC Prize Foundation published results for that preview on ARC-AGI-1, a puzzle benchmark designed to test novel pattern reasoning. o3-mini, a smaller and cheaper model aimed at science, math and coding, launched on 31 January 2025 and became the first reasoning model offered to free ChatGPT users. The full o3 and the smaller o4-mini arrived on 16 April 2025, and o3-pro followed for Pro subscribers in June 2025. OpenAI has since moved on: its posts now note that o1 is deprecated and that o3 and o4-mini are no longer available in ChatGPT and are deprecated in the API, with newer GPT-series reasoning models taking their place.

The published numbers explain the attention. On the 2024 AIME, a qualifying exam for the USA Mathematical Olympiad, OpenAI reported GPT-4o solved 12% of problems on average, while o1 averaged 74% with a single attempt, 83% with a majority vote over 64 samples, and 93% when re-ranking 1,000 samples with a learned scoring function. On Codeforces programming contests o1 reached the 89th percentile, and on GPQA Diamond, a set of graduate-level science questions, OpenAI said o1 was the first model to exceed the PhD experts it recruited for comparison. For o3, OpenAI reported 20% fewer major errors than o1 on difficult real-world tasks in external expert evaluations, and the ARC Prize Foundation reported the December 2024 preview scored 75.7% on its semi-private ARC-AGI-1 set at about $26 per task, and 87.5% using roughly 172 times more compute.

o3 also changed how reasoning connects to tools. OpenAI said o3 and o4-mini were trained with RL to decide when and how to use tools such as web search, a Python interpreter and image generation, and could chain several tool calls inside one reasoning process. They were also the first OpenAI models to put images into the chain of thought, cropping or rotating a picture mid-reasoning. With a Python interpreter, OpenAI reported o4-mini reached 99.5% on AIME 2025 on a single attempt, while cautioning that tool-assisted scores should not be compared with scores without tools.

Compared with a standard chat model, a reasoning model trades speed and cost for accuracy on problems with checkable answers. OpenAI's own human preference tests found o1-preview strongly preferred for data analysis, coding and math but not preferred on some natural language tasks such as writing. Compared with open alternatives, DeepSeek-R1, released in January 2025 and later published in Nature, showed that similar reasoning behavior, including self-reflection and verification, could emerge from reinforcement learning without human-written reasoning examples, and its authors described the training method publicly, whereas OpenAI has not published the o-series architecture or training recipe.

The limitations are well documented. The raw chain of thought is hidden: OpenAI chose to show users only a model-written summary, citing user experience, competitive advantage and the goal of monitoring the unaltered reasoning for problems such as manipulation. More thinking does not guarantee more truth. TechCrunch, citing OpenAI's o3 system card, reported that on OpenAI's PersonQA test of facts about people o3 hallucinated in 33% of answers and o4-mini in 48%, against 16% for o1, and the nonprofit lab Transluce found o3 sometimes claimed to have run code it could not have run. Benchmark context matters too: the ARC Prize Foundation noted that the o3 it tested had been trained on part of the ARC public training set and was not the same version later released.

In practice, reasoning models cost more per answer than chat models because every hidden reasoning token is billed. OpenAI's documentation recommends reserving at least 25,000 tokens for reasoning and output when starting out, warns that a response can run out of room before any visible text appears, and offers a reasoning effort setting and a cap on output tokens to control spending. Latency ranges from seconds to minutes. For simple lookups or casual writing, a faster non-reasoning model is usually the better fit.