OpenAI Explained: From GPT-3 to GPT-5

OpenAI Explained in one paragraph: OpenAI is the company behind ChatGPT and the GPT family of language models. GPT-3, described in a 2020 research paper, showed that a very large language model could handle many new tasks from just a few examples in its prompt. GPT-3.5 then made that ability conversational, and ChatGPT, launched on November 30, 2022, put it in front of ordinary users for the first time. GPT-4 added image understanding and stronger reasoning, GPT-4o made the model faster and able to work across text, image, and audio, the o-series taught models to think step by step before answering, and GPT-5, launched on August 7, 2025, combined those threads into one unified system. Checked 1 November 2026.

If you have ever wondered why ChatGPT sometimes feels like a helpful expert and sometimes confidently gets something wrong, the answer is mostly in this history. Each GPT generation changed something specific: how much the model could learn from examples, how well it followed instructions, what kinds of input it could accept, how long it would reason before answering, and how carefully it was trained to stay helpful without overstepping. This article follows that chain, generation by generation, in plain language, and explains what each step actually changed for the person typing the prompt.

A note on scope before we start: this article follows the GPT line from GPT-3 to GPT-5, because that is the arc most people mean when they ask how ChatGPT got so capable so quickly. OpenAI has continued to update the GPT-5 family since August 2025, and model names and availability in ChatGPT and the API change over time, so treat any statement about which model is the current default as time-sensitive and check OpenAI’s own model pages before making a decision that depends on it. Where OpenAI has not published a figure — for example, the parameter counts of its current frontier models — this article says so rather than repeating a rumor as fact.

1. What OpenAI Is, and What It Is Not

<img src="

” alt=”A modern research building at dawn with soft light, suggesting a laboratory where AI research becomes everyday products” />

OpenAI began as an AI research organization, publicly introduced in December 2015, with a stated mission to make sure powerful AI benefits humanity broadly. That framing matters, because it explains why the company’s early output looked like research papers and demos rather than consumer products. The organization spent its first years publishing work on language models, robotics, and reinforcement learning, and building the computing and engineering base that later made very large models possible. It was not, in those early years, a household name, and it did not yet have a mass-market product.

It is easy to confuse OpenAI with ChatGPT, but they are not the same thing. OpenAI is the organization; ChatGPT is one of its products, a conversational interface built on top of GPT models. Other products and services use the same underlying models through an application programming interface (API), which lets developers build the models into their own apps, and through partnerships such as Microsoft’s integration of OpenAI models into its products. When people say “GPT” they usually mean the model family; when they say “ChatGPT” they mean the chat product that most people actually use.

OpenAI’s structure has also changed over time, and its governance has at times been the subject of intense public attention, including the brief leadership turmoil of November 2023. This article does not try to settle questions about corporate structure, because those details matter less to a reader who wants to understand the models themselves. What matters for understanding GPT-3 through GPT-5 is the research thread: a bet, made early and pursued consistently, that training a single general-purpose language model at larger and larger scale would produce capabilities that narrower, task-specific systems could not match.

That bet is the through-line of everything that follows. Each generation tested it in a new way — bigger models, better instruction-following, new input types, and eventually models that reason before they answer. None of those steps made the models infallible, and none of them turned a language model into a person. What they did, cumulatively, was turn a research curiosity into a tool that hundreds of millions of people use for writing, coding, learning, and everyday problem-solving.

2. What GPT Means, and How a GPT Model Actually Works

GPT stands for Generative Pre-trained Transformer. Each word does real work. “Generative” means the model produces new text rather than just classifying or retrieving existing text. “Pre-trained” means it first learns from a very large body of text, before it is adapted for particular uses. “Transformer” is the name of the neural-network architecture, introduced in a 2017 paper by researchers at Google, that lets a model weigh the relationships between words in a passage — which earlier words matter most for understanding the current one — far more effectively than previous approaches.

In practice, a GPT model works by predicting the next piece of text, called a token, given everything that has come before it. A token is roughly a word or part of a word; the sentence you are reading right now would be broken into dozens of tokens before a model processed it. During pre-training, the model sees enormous amounts of text and repeatedly adjusts its internal settings so that its predictions get better. It is not memorizing answers, and it is not looking things up in a database. It is learning statistical patterns: how language is structured, how facts tend to be expressed, how arguments are built, how code is organized, and how different styles of writing differ from one another.

That simple mechanism explains both the power and the limits of these models. Because the model has absorbed patterns from so much text, it can continue a story, summarize a document, translate a paragraph, or draft an email in a plausible style, often without any task-specific training. But because it is predicting likely text rather than consulting a verified source, it can also produce fluent, confident statements that are simply wrong. That failure mode has a name — hallucination — and it recurs in every generation discussed here, even as each generation reduces it. If you remember only one technical idea from this article, remember this: a GPT model is a prediction machine, not a knowledge database, and every good use of it starts from that distinction. Our companion guide to how AI works, from transformers to next-token prediction, explains the underlying mechanism in more depth.

The other idea worth fixing early is the difference between a model and a product. The same underlying model can behave very differently depending on how it is prompted, what tools it can use, and how it was fine-tuned after pre-training. Much of the progress from GPT-3 to GPT-5 was not just about making the base model bigger. It was about everything wrapped around it: instruction tuning, safety training, tool use, and the ability to reason before answering. Those wrappings are what turned a raw text predictor into ChatGPT.

3. GPT-1 and GPT-2: Small Beginnings That Proved the Idea

Before GPT-3, there were two earlier models that most people have never used but that set the pattern. GPT-1 appeared in 2018, and GPT-2 followed in 2019. By today’s standards both were small — GPT-2 had 1.5 billion parameters, a parameter being one of the adjustable values the model learns during training. GPT-2 was notable at the time because OpenAI initially withheld its full release over concerns that it could be misused to generate misleading text, a decision that now reads as an early sign of the safety debates that would follow every later release.

What GPT-1 and GPT-2 proved was that pre-training followed by fine-tuning could work. Instead of building a separate system for every task — one for translation, one for summarization, one for question answering — you could pre-train one model on a large amount of text and then adapt it, with relatively little extra training, to many tasks. GPT-2 could already generate paragraphs of coherent text, even if it lost the thread over longer passages and made obvious mistakes. The results were impressive enough to justify scaling up, and that decision — to go much, much bigger — is what produced GPT-3.

It is worth pausing on how modest this starting point was, because it corrects a common myth: that ChatGPT appeared out of nowhere in 2022. It did not. It was the visible tip of a research program that had been running, in public, for years, with each model building directly on the last. The people who followed the field in 2019 could see GPT-2’s fluent but unreliable text and reasonably ask what would happen with a model a hundred times larger. GPT-3 was the answer to that question.

4. GPT-3: The Few-Shot Breakthrough of 2020

GPT-3 was described in May 2020 in the paper “Language Models are Few-Shot Learners,” by Tom Brown and colleagues at OpenAI. Its headline numbers are historical and well documented: the largest version had 175 billion parameters, compared with GPT-2’s 1.5 billion, and the models in the paper were trained on 300 billion tokens drawn from sources including filtered web text, books, and Wikipedia. Those figures are worth stating precisely, because they are often misquoted, and because they mark the moment when scale itself became the story.

The paper’s central finding was not just that the model was bigger, but that bigness changed how it could be used. Earlier models typically needed fine-tuning — extra training on examples of a specific task — before they could do that task well. GPT-3 could often do a new task with no fine-tuning at all, simply by being shown the task in its prompt: an instruction, plus zero, one, or a few worked examples. The researchers called this zero-shot, one-shot, and few-shot learning, and the umbrella term for it is in-context learning. Give GPT-3 a few examples of English-to-French translation in the prompt, and it would often translate a new sentence correctly. Give it examples of question-and-answer pairs, and it would often answer a new question in the same format. The model had not been retrained; it was inferring the task from context.

That discovery mattered for a practical reason: it made the model general-purpose in a way that felt new. A developer no longer always needed a labeled dataset and a training run to use the model for a new job. A well-written prompt could be enough. OpenAI made GPT-3 available through an API, and a wave of startups and experiments followed — writing assistants, summarizers, code helpers, and chatbots, many of them built on the same model with different prompts wrapped around it. GPT-3 was still a raw model, though, not a product for ordinary users. Using it meant working through the API or a developer playground, writing prompts in a format that felt technical, and tolerating its tendency to ramble, repeat itself, or state falsehoods with confidence. The paper itself was candid about those limits, including loss of coherence over long passages and errors in basic physical common sense.

GPT-3 also established a pattern that would repeat: each new capability brought a new set of worries. If a model could write news-style text that human evaluators struggled to distinguish from human writing — another finding reported in the paper — then it could also be used to produce misleading content at scale. That tension between capability and misuse has accompanied every GPT release since, and it is one reason OpenAI’s later models arrived wrapped in more safety training, usage policies, and product controls than GPT-3 ever had.

5. InstructGPT and GPT-3.5: Learning to Follow Instructions

GPT-3’s biggest practical weakness was not its knowledge but its manners. A raw language model continues text; it does not necessarily do what you asked. Ask it a question and it might answer, or it might invent a follow-up question, or continue in the style of a quiz, or drift into something else entirely. For a research demo that is interesting. For a product, it is unusable. The fix came from a line of work published in early 2022 as InstructGPT, which fine-tuned GPT-3 using reinforcement learning from human feedback, usually shortened to RLHF.

RLHF works roughly like this. Human trainers rank different model responses to the same prompt, preferring the ones that follow the instruction helpfully, truthfully, and safely. The model is then trained to favor responses like the preferred ones. A second stage uses a reward model — itself trained on those human rankings — to further steer the model toward preferred behavior. The result is a model that still predicts text, but that has been shaped to behave more like an assistant: it answers the question asked, follows the format requested, and declines or hedges more appropriately when it should not comply.

The models that came out of this work were grouped under the name GPT-3.5. That label covers a family of improved GPT-3 models rather than a single model, and the distinction matters because people sometimes talk about “GPT-3.5” as if it were one clean release. It was not; it was a series of instruction-tuned models, including text-davinci-003, that progressively made the raw power of GPT-3 usable through ordinary conversation. The most important of those models was the one fine-tuned for dialogue, because it became ChatGPT.

This step is easy to underrate beside the headline parameter counts, but for most users it was the step that changed everything. Scale made GPT-3 capable; instruction tuning made it cooperative. Without RLHF and its successors, ChatGPT would have felt like a clever autocomplete engine. With them, it felt like talking to something that was trying to help. Every later improvement — GPT-4, GPT-4o, the reasoning models, GPT-5 — kept and refined this alignment layer, even as the base models changed underneath it.

6. ChatGPT: November 30, 2022, and the Moment AI Went Mainstream

On November 30, 2022, OpenAI released ChatGPT as a free research preview, built on a model from the GPT-3.5 series. The announcement was modest — a blog post describing a model trained to interact conversationally, answer follow-up questions, admit mistakes, challenge incorrect premises, and reject inappropriate requests. There was no grand product launch, no advertising campaign, and, by OpenAI’s own later account, no expectation of what followed. Within days, millions of people were trying it, sharing screenshots of essays, poems, code, and jokes, and discovering both its fluency and its flaws in public, together.

Why did ChatGPT take off when GPT-3, which was arguably just as capable in raw terms, had not? The answer is mostly about form. GPT-3 required a developer’s mindset: you wrote prompts and tuned settings. ChatGPT required only a question. The dialogue format hid the machinery — the tokens, the prompts, the parameters — behind a text box that anyone could use. It also remembered context within a conversation, so a follow-up question made sense without repeating everything. For the first time, the average person could experience a large language model without any technical mediation, and the experience was good enough, and strange enough, to share.

The launch also made the model’s weaknesses public in a way that research papers never had. ChatGPT confidently invented facts, cited sources that did not exist, and produced plausible-sounding but wrong answers, especially on topics where a confident tone masked thin knowledge. OpenAI’s launch post warned about exactly this, noting that the model could produce plausible-sounding but incorrect or nonsensical answers. Teachers, employers, and journalists began debating, almost overnight, what it meant for homework, hiring, and trust in text. Those debates have never really stopped; they have simply moved on to each new model.

Commercially, ChatGPT changed OpenAI itself. A research organization became, within months, a consumer technology company with a subscription product — ChatGPT Plus, introduced in February 2023 — and an API business serving developers at scale. That shift funded the enormous computing costs of training larger models, but it also changed the incentives around each release: models were now judged not just by benchmark scores but by how they felt to use, how quickly they answered, and how many people kept coming back. The rest of this story is best understood as OpenAI trying to keep improving the model while serving that mass audience. The rapid growth of ChatGPT as a way people find answers has also reshaped the web itself, as shown by recent data on how much traffic ChatGPT actually sends to sites compared with Google search.

7. GPT-4: Multimodal, More Reliable, and Much Harder to Fool

On March 14, 2023, just over three months after ChatGPT launched, OpenAI announced GPT-4. It was the first GPT model that was multimodal in a meaningful sense: it could accept image inputs as well as text, and produce text outputs. At launch, image input was limited — text input was available through ChatGPT and the API, while image input rolled out more gradually — but the direction was clear. The model was no longer confined to words alone; it could look at a picture, a chart, or a diagram and reason about it.

OpenAI described GPT-4 as more reliable, creative, and able to handle more nuanced instructions than GPT-3.5, and it backed that claim with a striking comparison: on a simulated bar exam, GPT-4 scored around the top 10 percent of test takers, while GPT-3.5 had scored around the bottom 10 percent. That comparison was widely quoted, and it is worth understanding what it does and does not show. It shows that GPT-4 handled complex, multi-step professional reasoning far better than its predecessor. It does not show that GPT-4 could practice law, or that exam performance equals real-world judgment. Benchmarks are useful signals, not certifications, and OpenAI itself warned at the time that GPT-4 still had limitations, including a tendency to hallucinate, and that it was, in the company’s words, still flawed and still limited.

Two other GPT-4 details matter for the story. First, OpenAI did not publish GPT-4’s parameter count or training details, breaking with the openness of the GPT-3 paper. From GPT-4 onward, the exact size of frontier models became undisclosed, partly for competitive reasons, and that is why this article gives parameter counts only for GPT-3 and earlier models. Anyone who quotes a precise GPT-4 or GPT-5 parameter count is repeating a rumor or an estimate, not an OpenAI disclosure. Second, GPT-4’s training included six months of additional alignment work, using lessons from adversarial testing and from ChatGPT itself, aimed at improving factuality, steerability, and refusal behavior. That work continued the InstructGPT thread: the model was not just bigger, it was more carefully taught how to behave.

For users, GPT-4’s arrival created the first real tiering in ChatGPT: free users stayed on the GPT-3.5-era model, while paying subscribers got GPT-4. That split, in various forms, has persisted ever since, and it is the origin of a question users still ask today: is the paid model worth it for what I actually do? The honest answer depends on the task. For short, simple questions, the difference was often small. For long documents, tricky reasoning, code, and professional work, GPT-4’s advantage was large and obvious.

8. GPT-4 Turbo and GPT-4o: Faster, Cheaper, and Able to Hear and See

After GPT-4, OpenAI’s releases split into two kinds: more capable models, and models optimized to be faster and cheaper to run. GPT-4 Turbo, announced in late 2023, belonged to the second kind. It offered a much longer context window — the amount of text the model can consider at once — and lower running costs for developers, which made it practical to build applications that fed the model long documents or large codebases. For users, the practical effect was less visible than GPT-4 itself, but for developers it mattered enormously: a model that can hold a whole report in context is useful in a different way from one that can only see a few pages.

GPT-4o, announced on May 13, 2024, was the bigger public step. The “o” stands for “omni,” and the idea was a single model trained end to end across text, vision, and audio, so that all inputs and outputs were processed by the same neural network rather than bolted together from separate systems. In practice, that meant GPT-4o could reason across audio, vision, and text in real time, and respond to spoken input with a speed — OpenAI reported responses to audio inputs in as little as 232 milliseconds, averaging around 320 milliseconds — that felt close to human conversational timing. Text and image input rolled out in ChatGPT and the API at launch, with voice and video capabilities following.

GPT-4o also changed the economics of access. OpenAI made its text and image capabilities available to free ChatGPT users, with paid users getting higher usage limits, and it described the model as twice as fast and half the price of GPT-4 Turbo in the API. That combination — flagship-level intelligence available for free, with speed as a feature in its own right — is why GPT-4o became the default ChatGPT experience for most people through much of 2024 and into 2025. It also set the stage for voice mode, in which talking to ChatGPT began to feel less like dictating to a machine and more like a conversation, with interruptions, tone, and pacing that earlier text-only models could not offer.

There is a lesson in the GPT-4o era that is easy to miss: progress stopped being only about raw intelligence. Speed, cost, latency, and multimodality became first-class improvements. A model that answers in a second is useful in situations where a model that takes thirty seconds is not, even if the slower model is slightly smarter. GPT-4o’s success came from getting that balance right, and it foreshadowed GPT-5’s design, which is organized around choosing the right amount of effort for each question rather than always running at full power.

9. The o-Series: Teaching Models to Think Before They Answer

On September 12, 2024, OpenAI introduced a new line that broke the GPT numbering: o1, a series of reasoning models designed to spend time thinking before responding. The idea was simple to state and hard to build. Instead of answering immediately, the model generates an internal chain of thought — working through a problem step by step — and only then produces its final answer. That extra thinking time, sometimes called test-time compute, trades speed for accuracy on problems where careful reasoning matters: mathematics, science, coding, and multi-step logic.

The early results were dramatic on the kinds of problems that reward step-by-step work. OpenAI reported that o1’s performance on a challenging mathematics competition, the American Invitational Mathematics Examination, was far above GPT-4o’s, and similar gaps appeared on science and coding evaluations. But o1 was also, by design, a specialist. The preview release lacked some features that made ChatGPT broadly useful, such as web browsing and file uploads, and for many everyday tasks GPT-4o remained the better choice. OpenAI was explicit about that trade-off at launch: reasoning models were not a replacement for GPT-4o, but a new tool for a different kind of problem.

The o-series then evolved quickly. A full version of o1 arrived in December 2024, alongside a smaller, faster o1-mini, and in 2025 OpenAI released o3 and o4-mini, with o3 positioned as its most capable reasoning model to date and o4-mini as a faster, more cost-efficient option. The naming skipped o2, reportedly to avoid confusion with a telecom brand. What matters more than the version numbers is the shift in thinking they represent: OpenAI had found a second axis of improvement beyond pre-training scale. If pre-training made models know more, reasoning made them use what they knew more carefully. That second axis is the key to understanding GPT-5, which is essentially the GPT line and the o-line merged into one system.

For ordinary users, the o-series introduced a new habit: asking the model to show its working, or to take its time on a hard problem, actually helped. Prompts like “think step by step” had always helped to some degree, but with reasoning models the thinking was built into the system rather than coaxed out by clever prompting. The cost was patience — reasoning models take longer — and the benefit was reliability on exactly the tasks where a quick, fluent answer is most dangerous: calculations, proofs, and plans where one wrong step ruins everything after it.

10. GPT-4.5 and the Road to GPT-5: Scale Meets Reasoning

On February 27, 2025, OpenAI released GPT-4.5, describing it as its largest and most knowledgeable model yet, and its best model for chat. GPT-4.5 was deliberately not a reasoning model: it was built by scaling up unsupervised learning — the pre-training stage — and refining it with supervised fine-tuning and human feedback, and OpenAI presented it as better at recognizing patterns, drawing connections, and generating creative insights, with a more natural feel in conversation. Its strength was intuitive, broad knowledge rather than step-by-step problem-solving, and OpenAI was candid that reasoning models outperformed it on coding and mathematics.

GPT-4.5 matters to this story because it clarifies what GPT-5 was trying to combine. By early 2025, OpenAI had two strong but different kinds of model. One kind, exemplified by GPT-4o and GPT-4.5, was fast, fluent, broadly knowledgeable, and good at conversation. The other kind, exemplified by o1 and o3, was slower, more deliberate, and much stronger on hard reasoning. Users had to choose between them, or OpenAI had to choose for them, and neither option was ideal. A user asking a simple question did not need a reasoning model; a user working through a hard problem did not want a fast model that guessed.

The releases of spring 2025 — o3 and o4-mini in April, and the continued refinement of the GPT-4 line — were, in retrospect, the last steps before the merger. They also showed OpenAI’s growing emphasis on agentic capabilities: models that could use tools, browse, run code, and carry out multi-step tasks with less hand-holding. That emphasis, combined with the reasoning work and the multimodal work, set the design brief for GPT-5. It would not just be a bigger GPT-4. It would be a system that could decide, question by question, whether to answer quickly or think deeply, and that could use tools along the way.

It is also worth noting what GPT-4.5 did not do: it did not settle the question of whether scale alone was running out of steam. OpenAI’s own framing — presenting GPT-4.5 as a step in scaling, while pointing to reasoning as the frontier — suggested that the company saw the future as a combination of both. GPT-5 would test that view directly, and its reception would show how much users valued a unified system over a menu of specialized models.

11. GPT-5: One Unified System, Launched August 7, 2025

On August 7, 2025, OpenAI introduced GPT-5, describing it as its best AI system yet and a significant leap in intelligence over all its previous models, with state-of-the-art performance across coding, math, writing, health, and visual perception. The launch mattered less for any single benchmark than for its architecture. GPT-5 in ChatGPT is a unified system: a smart, efficient model that answers most questions, a deeper reasoning model called GPT-5 thinking for harder problems, and a real-time router that quickly decides which to use based on the conversation type, complexity, tool needs, and the user’s explicit intent — for example, if you tell it to think hard about a problem. Once usage limits are reached, a mini version handles remaining queries.

That router is the idea that ties the whole GPT-3 to GPT-5 story together. GPT-3 asked: can one model do many tasks from examples? GPT-3.5 asked: can it follow instructions? GPT-4 asked: can it see as well as read? GPT-4o asked: can it hear, see, and respond in real time? The o-series asked: can it reason before answering? GPT-5 asks: can one system do all of that, and choose the right mode for each question, without making the user choose? For users, the practical promise was that they would no longer need to know which model to pick. They could just ask, and the system would decide how much thinking the question deserved.

OpenAI’s launch claims for GPT-5 were specific, and they are worth quoting with attribution rather than as settled fact. The company reported 94.6 percent on AIME 2025 mathematics without tools, 74.9 percent on SWE-bench Verified for real-world coding tasks, 84.2 percent on MMMU for multimodal understanding, and 46.2 percent on HealthBench Hard. It also reported that, with web search enabled on prompts representative of ChatGPT traffic, GPT-5’s responses were about 45 percent less likely to contain a factual error than GPT-4o’s, and that GPT-5 thinking was about 80 percent less likely to contain a factual error than o3’s. Those are OpenAI’s own evaluations, on its own benchmarks and methodology, and independent verification matters — but the direction is consistent with the story of this article: each generation reduced, without eliminating, the hallucination problem that has shadowed the GPT family since GPT-3.

GPT-5 also arrived with a changed tone. OpenAI said it was less effusively agreeable than GPT-4o, used fewer unnecessary emojis, and was more subtle and thoughtful in follow-ups, after an earlier GPT-4o update had unintentionally made the model overly sycophantic and had to be rolled back. New preset personalities were offered as a research preview, letting users set how ChatGPT interacted without writing custom prompts. And GPT-5 was made available to all ChatGPT users, with Plus subscribers getting more usage and Pro subscribers getting GPT-5 Pro, a version with extended reasoning for the most demanding tasks. The launch was not without friction — early users complained about the removal of older models and the feel of the new system, and OpenAI restored access to some legacy models — but the unified-system design survived, and it is the design that later GPT-5 updates built on. Readers comparing ChatGPT with its closest rival may also want our Claude vs ChatGPT complete comparison for the other side of that choice.

12. GPT-5 for Developers: Mini, Nano, and Control Over Thinking

For developers, GPT-5 arrived in a different form from ChatGPT, and understanding that difference clears up a lot of confusion. In the API, OpenAI released GPT-5 in three sizes — gpt-5, gpt-5-mini, and gpt-5-nano — to let developers trade off performance, cost, and latency. The API version of GPT-5 is the reasoning model that powers maximum performance in ChatGPT, while the non-reasoning model used in ChatGPT is available in the API as gpt-5-chat-latest. That split is deliberate: an app that needs the strongest reasoning will choose differently from an app that needs fast, cheap responses at scale.

Two new API controls also mattered. The verbosity parameter, with values low, medium, and high, lets developers steer how long answers should be by default, though an explicit instruction in the prompt still takes precedence. The reasoning_effort parameter, which already existed with low, medium, and high values, gained a minimal value for GPT-5, letting developers get answers back faster without extensive reasoning first. Together, those controls embody the GPT-5 philosophy in developer form: thinking is a resource to be spent wisely, not a switch that is always on or always off. OpenAI also added custom tools, letting GPT-5 call tools with plain text instead of only JSON, with developer-supplied grammars to constrain the format.

The scale of context grew, too. In the API, all GPT-5 models could accept a maximum of 272,000 input tokens and emit up to 128,000 reasoning and output tokens, for a total context length of 400,000 tokens. In plain terms, that is enough to hold very long documents, large codebases, or extended conversations in context at once. That capability, combined with improved tool calling — OpenAI reported 96.7 percent on the tau-bench telecom tool-calling benchmark — is what made GPT-5 practical for agentic work: software that does not just answer questions but carries out multi-step tasks, calling tools, checking results, and adapting when something goes wrong. Enterprise adoption followed that capability, with ChatGPT put to work inside large workplace systems, as in OpenAI putting ChatGPT to work inside ServiceNow workflows.

None of this means a developer should pick the biggest model by default. The sensible rule, and the one OpenAI’s own design encourages, is to match the model and the reasoning effort to the task: mini or nano for high-volume, straightforward work; the full model for hard reasoning; minimal reasoning effort when speed matters and the task is simple; higher effort when correctness matters more than latency. That decision — not the model name alone — is where most of the real cost and quality trade-offs in building with GPT-5 now live.

13. Using GPT Models Well: What Five Generations Teach About Prompting

Five generations of GPT models teach a consistent lesson about prompting: the model rewards clarity, context, and checking. The earliest lesson, from GPT-3’s few-shot learning, still holds: showing the model what you want, with an example or two, works better than describing it abstractly. If you want a summary in a particular format, show it one summary in that format. If you want code in a particular style, show it a snippet. The model is a pattern-completer, and a good prompt gives it the pattern to complete.

The second lesson, from the instruction-tuning era, is to say what you mean about role, audience, format, and constraints, all in the same prompt. “Write a report” is a weak prompt. “Write a 500-word report for a non-technical manager, with three sections, a recommendation at the end, and no jargon” is a strong one, because it removes the ambiguity the model would otherwise have to guess at. GPT-5’s improved instruction-following makes this more effective, not less necessary: a model that follows instructions faithfully will follow vague instructions faithfully, and produce a vague answer.

The third lesson, from the reasoning era, is to match effort to difficulty. For a simple factual question, a quick answer is fine. For a calculation, a plan, or a decision with real consequences, ask the model to work step by step, show its reasoning, and flag its assumptions — and then check the steps yourself. The fourth lesson, from the multimodal era, is to give the model the right input: a photo of the error message, the relevant page of the document, or the table it needs to analyze, rather than a description of it from memory. And the final lesson, from every generation, is to verify anything that matters. Use the model to draft, brainstorm, explain, and explore; use a primary source to confirm facts, figures, and quotations before you publish or act on them.

It is also worth being honest about what prompting cannot fix. No prompt turns a language model into a reliable source of truth, a licensed professional, or a substitute for judgment. A well-prompted model can still hallucinate, can still misunderstand a subtle instruction, and can still produce a confident answer to a question it should have refused or hedged. The generations from GPT-3 to GPT-5 narrowed those failure modes considerably. They did not remove them, and any guide that promises otherwise is selling something.

14. Limits, Safety, and What GPT-5 Is Not

GPT-5 is not artificial general intelligence, and OpenAI has not claimed that it is. It is a very capable language model system: it predicts text, reasons through problems when asked to think, uses tools, and handles images and audio. It does not understand the world the way a person does, it does not have beliefs or intentions in the human sense, and it does not know when it is wrong unless its training and its tools give it a way to check. Calling it “PhD-level” — a phrase OpenAI’s chief executive used at the GPT-5 launch to describe the feel of talking to it — is a metaphor about fluency and breadth, not a credential. A model can write like an expert on a topic and still make an elementary error in the next sentence.

Safety training has evolved alongside capability, and GPT-5’s approach is worth understanding because it affects what users actually experience. Earlier ChatGPT safety training relied heavily on refusal: for a disallowed request, refuse. GPT-5 introduced what OpenAI calls safe completions, which aim to give the most helpful answer possible while staying within safety boundaries — sometimes answering partially, or at a high level, and explaining a refusal transparently with safe alternatives where appropriate. The idea is to handle the large middle ground of dual-use questions, where intent is unclear or information could be used benignly or harmfully, with more nuance than a simple yes or no. OpenAI also rewrote its public guidance for how its models should behave; our coverage of OpenAI rewriting its AI rulebook around five values explains that framework and why it matters to everyday users.

There are also practical limits that have nothing to do with safety policy. GPT models have knowledge cutoffs — they were trained on data up to a point in time — and they can be out of date on fast-moving facts unless they can browse the web or use a tool. They can be misled by false premises in a prompt, sycophantic in agreeing with a user, or inconsistent across two conversations about the same topic. They can leak nothing and still reveal, through confident phrasing, an answer that was never verified. And their outputs reflect the biases and errors present in their training data, even after extensive fine-tuning. None of those limits is a reason to avoid using them; all of them are reasons to use them with a habit of verification.

The most useful way to think about GPT-5, then, is as the latest point on a curve that started with GPT-3’s few-shot surprise. GPT-3 showed that scale could produce generality. GPT-3.5 showed that generality could be made conversational. GPT-4 showed that it could see. GPT-4o showed that it could listen and respond in real time. The o-series showed that it could think before answering. GPT-5 showed that all of those abilities could live in one system that chooses how hard to think. The next step, whatever it is called, will probably not be a single breakthrough but another combination — more reasoning, better tools, longer memory, and more reliable grounding in sources. If the last five years are any guide, the most important question will not be how big the next model is, but how well it helps ordinary people think, create, and decide — and how honestly it says so when it does not know. Checked 1 November 2026.