OpenAI o-series – reasoning models lineage explained

OpenAI's o-series is the company's line of reasoning models: systems built to spend compute thinking through a problem before they answer, rather than producing the most fluent next token as fast as possible. The line began with o1-preview in September 2024, moved through o1-mini, o3 and o3-mini, and reached o4-class models in 2025. What separates these from the GPT lineage is not a bigger context window or a nicer chat manner. It is a training recipe – reinforcement learning applied to chain-of-thought traces – that teaches the model to reason in a hidden scratchpad, trading latency and token cost for measurably higher accuracy on hard mathematical, scientific and programming problems. That trade-off is the whole story, and it is what this analysis is about: how the lineage evolved, what "reasoning" actually buys you, and when it is worth paying for.

Because model versions, prices and benchmark scores in this field change month to month, treat every specific figure below as accurate as of publication and verify current details against OpenAI's own model pages before you commit budget or architecture to them.

The o-series in one paragraph

The o-series is OpenAI's reasoning-first model family, launched with o1-preview in September 2024 and designed for "think before answering" workloads. These models are trained with reinforcement learning over chain-of-thought reasoning, which means they learn to generate long internal deliberations – reasoning tokens – before emitting a final answer. The lineage, as of publication, runs o1 (and the cheaper o1-mini), then o3 (with o3-mini), then the o4-class models that represent the current frontier of the line. Each generation pushed harder on the same axis: more effective use of inference-time reasoning to solve problems that defeat standard chat models. The reason the family exists as a separate branch from GPT-4o and its successors is that the skills it optimises for – sustained multi-step logic, competition mathematics, algorithmic coding – do not come for free from scaling a conventional language model. They come from teaching the model to search through its own reasoning.

What "reasoning model" actually means

The phrase "reasoning model" is doing real technical work, not marketing work. Two things distinguish the o-series from a GPT chat model, and it helps to keep them separate.

Train-time reinforcement learning on chain-of-thought

Standard language models learn to predict the next token from enormous text corpora, then get aligned with human-preference tuning. You can coax them into showing their working by prompting "let's think step by step," and this inference-time scratchpad genuinely helps – but the model was never trained to be good at the deliberation itself. The o-series takes a different route. During training, OpenAI applies reinforcement learning to the model's chain-of-thought, rewarding reasoning traces that lead to correct answers. The model is not merely allowed to think out loud; it is optimised, over many iterations, to produce reasoning that actually resolves the problem. As OpenAI put it in its o1 announcement, the model "learns to recognise and correct its mistakes," break down hard steps and try alternative approaches when a line of attack fails. That capacity is baked in at train time, which is why you cannot simply prompt GPT-4o into being o1.

The hidden reasoning tokens

When you call an o-series model through the API, it generates a large volume of internal reasoning that you are billed for but do not see. OpenAI made an explicit decision to keep the raw chain-of-thought hidden, arguing that an unfiltered internal monologue is more useful for the model's own reasoning if it is not constrained by the safety and formatting pressures applied to user-facing output, and that exposing it raw carries safety and competitive risks. The practical consequence for builders is twofold: you pay for reasoning tokens you never read, and you lose the ability to audit exactly how the model reached its conclusion. You get a summary of the reasoning in some interfaces, but not the verbatim trace. This is a genuine trade-off, and one worth weighing against models that expose their thinking more openly.

The cost of thinking

The reasoning that makes o-series models accurate is also what makes them slow and expensive. A single hard query can consume thousands of hidden reasoning tokens before the visible answer begins. Latency climbs from the sub-second feel of a chat model into the range of seconds or, on the hardest problems, considerably longer. For a competition-mathematics problem this is a bargain. For "rewrite this email a bit more warmly," it is waste. The entire craft of using the o-series well is knowing which of those two cases you are in.

How it compares across vendors

OpenAI is not alone here. Anthropic's Claude models offer an extended thinking mode, in which the model produces a visible reasoning block before its answer, with the amount of thinking budget controllable by the caller – a more transparent approach than OpenAI's hidden traces, at the cost of exposing reasoning that can occasionally be messy. DeepSeek's R1, released as an open-weights reasoning model, made the approach reproducible and inspectable in a way the closed frontier labs did not, and demonstrated that RL-on-reasoning could be trained at a fraction of the assumed cost; we cover that lab in depth in our profile of DeepSeek, the open reasoning model lab. Google's Gemini line added its own "thinking" variants. The broader family resemblance, and the shared chain-of-thought heritage, is the subject of our explainer on reasoning models and chain-of-thought. The point for a builder is that "reasoning model" now names a category, not a single product, and the vendors differ meaningfully on transparency, price and control.

Capability picture per generation

The clearest way to read the lineage is as a sequence of jumps on hard reasoning benchmarks. A caution before the numbers: benchmark scores depend heavily on methodology, prompting, whether tools were allowed, and which specific model variant and reasoning-effort setting was used. The figures that follow are drawn from OpenAI's own announcements and independent evaluations current at publication, and they should be read as directional rather than definitive. For a grounding in what these tests actually measure, see our guide to AI benchmarks – MMLU, GPQA and the LMSYS Arena.

o1: the first credible RL-on-CoT model at the frontier

o1 was the proof of concept that mattered. On the American Invitational Mathematics Examination (AIME), o1 leapt far past GPT-4o – OpenAI reported it solving a large majority of problems where the conventional model solved only a handful. On GPQA Diamond, a set of graduate-level science questions, o1 reportedly exceeded the performance of human PhD experts in the relevant domains on that particular test. On competitive programming, it reached a Codeforces rating that placed it well into the upper percentiles of human competitors. None of these results made o1 a better chatbot. They demonstrated that train-time RL on reasoning traces produced a step-change on precisely the tasks where step-by-step correctness is everything. The companion o1-mini was positioned as a cheaper, faster reasoning model tuned especially for coding and STEM, sacrificing some breadth of world knowledge for a lower cost per token.

o3: the significant gain

o3, announced at the end of 2024 and rolled out through 2025, was the generation where the reasoning approach started to look like it might scale further than sceptics expected. The headline result came from the ARC-AGI benchmark, a test of abstract visual reasoning designed by François Chollet to be hard for systems that rely on memorised patterns. According to the ARC Prize team's analysis, o3 scored dramatically higher than any prior model on the ARC-AGI-1 evaluation – a result the organisers described as a genuine breakthrough, while being careful to note the enormous compute cost of the high-scoring configuration and cautioning that the benchmark is not a measure of general intelligence. o3 also posted large gains on FrontierMath, a set of research-level mathematics problems where earlier models scored close to nothing, and pushed AIME and Codeforces numbers higher still. The o3-mini variant brought much of the coding and math strength down to a lower price point with configurable reasoning effort.

o4: the current frontier

The o4-class models, current as of publication, continued the pattern and folded reasoning together with native tool use – the model can decide, mid-reasoning, to run code, search, or analyse an image, and weave the results back into its deliberation. This matters because a reasoning model that can act during its thinking closes much of the gap between "solving a problem on paper" and "getting a task done." Exact variant names, availability and scores shift, so confirm the current o4-tier lineup on OpenAI's model documentation rather than trusting any snapshot. The direction of travel is consistent: each generation gets more out of inference-time reasoning, and increasingly couples that reasoning to tools and agentic behaviour. That convergence with agent infrastructure is a theme we track in our coverage of AI agents.

When the o-series earns its cost

Reasoning models pay for themselves on a specific and identifiable class of problem: tasks where the answer is either right or wrong, where getting it right requires several dependent steps, and where a mistake early in the chain quietly poisons everything after it.

Competition and applied mathematics is the clearest case. When a problem needs you to set up equations, manipulate them correctly and check the result, the o-series' habit of working through and verifying its own steps produces answers a chat model frequently fumbles. Competitive and algorithmic programming is the second: not writing boilerplate, but the kind of problem where you have to find the right approach – a dynamic-programming formulation, the correct data structure, an edge case that breaks the naive solution. Scientific reasoning is a third, particularly the multi-step quantitative kind found in physics, chemistry and parts of biology. The protein-structure and molecular-modelling work we survey in our piece on 3D shapes and protein folding AI sits adjacent to the sort of reasoning these models are being pushed toward.

The unifying pattern is a long dependency chain with a verifiable endpoint. If you can tell whether the final answer is correct, and getting there needs sustained logic rather than recall or fluency, the o-series is where the extra cost buys real accuracy. Complex data analysis, difficult debugging, formal logic puzzles and legal or financial reasoning with many interacting constraints all fit the profile.

When not to use the o-series

The mirror image is just as important, and it is where builders waste the most money. For ordinary conversation – support chat, question-answering, routing, summarisation – a reasoning model is slower, pricier and no more accurate than GPT-4o or an equivalent chat model. For creative writing the o-series is often worse, not better: the deliberative training can make prose stiffer and more mechanical than the looser, more associative output of a standard model, and many writers report preferring 4o-class output for tone and voice.

Any real-time or low-latency requirement argues against reasoning models. If a user is waiting on a response in an interactive loop, the seconds a reasoning model spends thinking are seconds of dead air. High-volume, cost-sensitive workloads – classification at scale, bulk tagging, template filling – are the wrong place for a model that bills you for thousands of invisible reasoning tokens per call. And the simplest test of all: if 4o, Claude or an open model already succeeds reliably on your task, the o-series adds cost and latency for no measurable gain. Reasoning capacity is a tool for problems that are actually hard. Applied to easy ones, it is pure overhead.

API access and pricing

Access to the o-series has generally followed OpenAI's usage-tier system, with the most capable models sometimes gated behind higher spending tiers or verification at launch before broadening. Both ChatGPT subscribers and API developers have had routes to the models, though the specific variants exposed in each surface differ and change over time.

Pricing is the part builders most need to get right, and it is also the most time-sensitive, so treat the following as structural rather than exact and confirm against OpenAI's pricing page. Reasoning models are billed per million input and output tokens like other OpenAI models, but with a crucial wrinkle: the hidden reasoning tokens count as billed output. That means the real cost of an o-series call is often several times the cost of the visible answer alone, and it is why naive cost estimates based on the length of the response are badly wrong. Historically the full-size reasoning models (o1, then o3) have sat at a substantial premium over GPT-4o per token, while the mini variants (o1-mini, o3-mini) have offered much of the reasoning benefit at a fraction of the price. Configurable reasoning effort, introduced with the o3-mini generation, lets you dial the amount of thinking – and therefore the cost – up or down per request.

Rate limits scale with your usage tier, and reasoning models can hit them faster than you expect precisely because each call consumes so many tokens. If you are architecting a high-throughput system, plan for backoff and retry behaviour from the start; our guide to 429 Too Many Requests and OpenAI rate-limit errors covers the failure modes. For a broader view of where these models sit in the company's product stack, see our profile of OpenAI – the company, its models and its ecosystem.

Building with reasoning models

The most cost-effective production pattern is not "use the reasoning model" but "use the reasoning model for the hard step only." Route the bulk of a workflow through a fast chat model and hand off to the o-series exactly when a task requires it – the difficult algorithmic decision, the verification pass, the step where a wrong answer is expensive. A well-designed pipeline might use 4o to parse and structure a request, an o-series model to solve the genuinely hard sub-problem, and 4o again to format the result into readable output. You pay the reasoning premium on a small fraction of tokens rather than all of them.

Prompting reasoning models is different from prompting chat models, and some habits invert. The elaborate "think step by step, first do X, then do Y" scaffolding that helps a standard model can actively hurt a reasoning model, because it constrains a deliberation process the model already performs better on its own. OpenAI's own guidance has leaned toward keeping prompts clear and direct, stating the goal and the constraints, and letting the model find the path. Few-shot examples, which lift chat-model performance, sometimes degrade reasoning-model performance by anchoring it. The practical advice is to test both minimal and elaborate prompts on your actual task rather than assuming the chat-model playbook transfers.

Finally, weigh the visible-versus-hidden reasoning trade-off against your requirements. If you need an auditable reasoning trail – for regulated decisions, for debugging, for user trust – OpenAI's hidden chain-of-thought is a real limitation, and a model that exposes its thinking, whether Anthropic's Claude with extended thinking or an open-weights model like DeepSeek R1, may serve you better even at some cost in raw benchmark performance. Vendor neutrality is not a slogan here: the right choice depends on whether you value the highest score, the lowest latency, the most transparency or the lowest cost, and the o-series wins on some of those axes and loses on others.

Where the line is heading

The through-line from o1 to o4 is that inference-time reasoning, once it was trained rather than merely prompted, kept yielding returns generation after generation – and that those returns increasingly come bundled with the ability to act, not just to think. The open question is economic rather than technical. Reasoning that scores at the top of ARC-AGI can cost orders of magnitude more per task than a chat-model answer, and the field has not settled where the price of a "thought" lands as competition from open reasoning models and cheaper training methods bites. For builders, the durable lesson survives every version bump: match the model to the problem. The o-series is the sharpest instrument OpenAI makes for genuinely hard reasoning, and using it for anything else is paying frontier prices for a job a cheaper model already does well.