OpenAI launches o1, models that reason before answering

On September 12, 2024, OpenAI released o1-preview and o1-mini, models designed to spend more compute on multi-step problems before answering. The original announcement publishes results in mathematics, programming and science alongside product limits; those tests describe specific tasks rather than a general ability to reason without error.

The company had previewed this line of research under the codename Strawberry. Its first commercial result does not replace GPT-4o: it is designed for tasks where a fast answer matters less than a well-formed solution, such as a complex math problem, a scientific query, or debugging code.

Taking more time before answering

Conventional language models generate text by predicting, one word at a time, what the most likely continuation should be. o1 adds a longer internal reasoning phase. During training, OpenAI used reinforcement learning, a technique that rewards strategies leading to a correct answer and penalizes mistakes.

As a result, the model can try different approaches, detect when one does not work, and correct itself before delivering its final answer. That does not mean it understands a problem like a person or makes every response correct. It does change how it handles tasks that require chaining together several logical steps.

OpenAI does not show users the model's full chain of thought. Instead, the interface may display a brief indication of the process while the model works and then deliver the answer. That is an important precaution: internal reasoning should not be mistaken for a verifiable demonstration. In math, research, or professional work, the answer still needs human review and comparison against independent sources or calculations.

Better results on math and coding tests

The company accompanied the launch with results that capture o1's goal. On a qualifying exam for the International Mathematical Olympiad, GPT-4o correctly solved 13% of the questions, while o1 reached 83%, according to OpenAI. In Codeforces programming competitions, the model ranked in the 89th percentile.

These tests do not measure intelligence as a whole or guarantee the same performance in real-world work. But they highlight a known weakness of generative assistants: they can confidently explain an incorrect solution when a problem requires maintaining logical consistency across many steps. o1's advance lies precisely in reducing that failure in structured domains.

o1-preview is the new series' more capable general-purpose version. It is available today to ChatGPT Plus and Team subscribers, with an initial limit of 30 messages per week. It is also coming to the API for developers who meet OpenAI's access requirements.

The second version, o1-mini, focuses on programming and STEM subjects—the acronym for science, technology, engineering, and mathematics. OpenAI offers it as a faster alternative that is 80% cheaper than o1-preview: it costs $3 per million input tokens and $12 per million output tokens, compared with $15 and $60, respectively, for o1-preview. A token is a unit of text that models use to process and generate language.

More capability, but fewer features than GPT-4o

The launch has clear limitations. In ChatGPT, o1-preview and o1-mini still lack tools already common in GPT-4o, such as web browsing, file uploads, and image analysis. They are also slower: that pause is not an accidental flaw but the cost of devoting more resources to reasoning.

This means choosing a model based on the task. For drafting, summarizing a document, or keeping a conversation moving quickly, GPT-4o will remain the more practical option in many cases. For solving a difficult technical exercise, reviewing a programming strategy, or exploring a scientific hypothesis, o1 may justify the wait and the higher price.

OpenAI has also evaluated this family under its framework for advanced-risk preparedness. The company classifies o1-preview as presenting medium risk in capabilities related to chemical, biological, radiological, and nuclear threats, and applies specific safeguards to those queries. It is a sign that more capable reasoning does not just improve useful answers: it also requires raising the bar for safety evaluation before access is expanded.

The first version of o1 is therefore less a universal assistant than a bet on models that know when to stop and work through a problem. Its evolution will depend on whether OpenAI can extend that reasoning to more tools and tasks without turning every everyday query into a costly wait.

More inference time changes the economics

A model that explores several paths may improve on structured problems, but it consumes latency and resources differently. Compare accepted-answer cost rather than token price alone, including failed attempts, duration, review and external tools. For a routine request the wait may buy nothing; for a mathematical proof it may.

Selection starts by classifying the task. Writing, extraction and conversation often reward speed; planning, tested code and verifiable calculation may benefit from extra internal search. Sending everything to the most reflective model wastes budget and makes an interface unpredictable.

Private reasoning does not replace a proof

The final answer should expose what another person needs to check: equations, assumptions, executed tests or sources. A long private chain cannot be audited and does not guarantee a valid route. Evaluate the visible artefact and reproduce it with an independent tool.

For programming, compile, run tests and review the change instead of scoring its explanation. In science, verify units, references and boundary conditions. In planning, simulate constraints and preserve a safe outcome when information is missing.

How to read a benchmark jump

Identify the dataset, scoring rule, sampling opportunities and possible contamination. A strong result on an olympiad or competition demonstrates progress under that protocol; it says nothing by itself about business documents, languages or current facts. A local replication mixes solved and novel cases with a blind evaluator.

Failures deserve more weight than the average when an answer can execute code or support a sensitive decision. Record the distribution: how many cases are correct, how many admit uncertainty and how many offer a convincing false solution. An average improvement can be unacceptable if the last group grows.

The production route adds a time and budget limit. If the model does not finish within them, the system returns an explicit state rather than presenting a partial answer as final. It also preserves a fallback: another model, a deterministic tool or human review. Longer thinking must not block the process indefinitely.

Track improvement with longitudinal cases. Repeating the same suite after each version shows whether progress on hard problems preserves simple tasks, format and safety. A new exam high does not compensate for a silent regression in routine work.

Calibration matters as much as accuracy

Besides correcting answers, ask the system to identify assumptions and flag which part needs a tool. Compare that signal with actual failure. A model that recognises its limits can work with reviewers; one that remains confident when wrong requires stronger barriers even if its average score is higher.

The transferable skill is to route each task by verifiability, error cost and benefit from longer thinking, then require external proof for the answer. The framework captures reasoning models without anthropomorphising them or mistaking additional compute for certainty.