OpenAI

OpenAI is an American artificial intelligence company based in San Francisco. It was founded in December 2015 as a non-profit research laboratory, became the most visible company in generative AI after ChatGPT was released in November 2022, and completed a corporate reorganisation in October 2025 that placed its operating business inside a public benefit corporation controlled by a non-profit foundation 1 2.

Its products include a consumer assistant, a developer platform, image and video generation tools, and enterprise software. Its models — the GPT family and the o-series reasoning models — are among the systems against which rival laboratories benchmark their own 3.

The company matters beyond its products. It has been the site of an unusual governance experiment, a public argument about how training data may be used, and a sustained debate about whether commercial pressure is compatible with its founding safety commitments 4 5.

OpenAI was announced on 11 December 2015 by Sam Altman, Greg Brockman, Ilya Sutskever, Elon Musk, John Schulman and Wojciech Zaremba, among others. It began as a non-profit with an initial pledge of about a billion dollars, no product, and a mission to ensure that artificial general intelligence benefits all of humanity 6.

That structure did not survive contact with the cost of compute. In 2019 the laboratory created a capped-profit entity, OpenAI LP, which allowed investors to earn returns up to a fixed multiple while the non-profit retained control. Microsoft invested a billion dollars that year and became the preferred cloud provider; a further investment, reported at around ten billion dollars, followed in January 2023 6.

ChatGPT, released on 30 November 2022, turned a research organisation into a consumer company. It reached an estimated hundred million users within two months, a pace without precedent for a software product, and it forced the company to build capacity, moderation and billing at speed 6.

GPT-4 arrived in March 2023, the o-series reasoning models from late 2024, and GPT-5 in August 2025. Alongside them came image generation, speech recognition, video generation and a coding agent, each of which widened the distance between the laboratory's research agenda and its commercial surface 3.

The consumer product is ChatGPT, available on the web and as applications, with a free tier and paid subscriptions. It accepts text, images and files, can search the web, and runs tools on the user's behalf; higher tiers raise usage limits and unlock the most capable models 1.

The developer platform is the commercial centre of gravity. It sells access to models by the token, with tools for fine-tuning, retrieval, speech and real-time interaction, and a batch service for large jobs. Most businesses that use OpenAI models do so through this interface, or through Microsoft's Azure OpenAI Service, rather than through the chat application 1.

Around the core models sit narrower products: image generation, video generation, transcription, and a coding agent that works in a sandboxed environment and returns patches rather than suggestions. ChatGPT Team, Enterprise and Edu packages add administration, data controls and, in the enterprise case, a default promise that inputs are not used for training 1.

The breadth is deliberate. Each surface — chat, coding, images, voice — generates revenue and usage data, and each makes the underlying models harder to replace with a competitor's. Video generation, demonstrated in early 2024 and released as Sora at the end of that year, remains a smaller business than the text models, but it has become a standard demonstration of what the systems can do.

OpenAI's models are Transformer networks trained on large text corpora, then tuned with reinforcement learning from human feedback, or RLHF, a method in which people rank outputs and the model is trained to prefer the higher-ranked ones. GPT-3, in 2020, showed that scale alone produced useful general behaviour; InstructGPT, in 2022, showed that tuning on human preferences made such models usable as assistants 6.

GPT-4, released in March 2023, accepted images as well as text and was the first OpenAI model deployed with a detailed system card describing evaluations and risks. GPT-4o, in May 2024, combined text, audio and vision in a single model trained end to end, which cut latency and made voice conversation feel less like a pipeline of separate components 3.

The o-series, beginning with o1 in September 2024, changed the training recipe rather than the architecture. These models are trained with reinforcement learning to produce long internal reasoning before answering, and they spend more computation at inference time on hard problems. The approach improved mathematics and code at the cost of latency and price.

GPT-5, released in August 2025, combined a fast general model with a slower reasoning mode and a router that chooses between them. The company described it as a single system rather than a family; independent evaluations found smaller gains over its predecessor than earlier generations had delivered 6.

OpenAI's ownership is the most complicated in the industry. Until 2025 it was a non-profit that controlled a capped-profit subsidiary; investors held convertible interests with a capped return, and the board answered to the mission rather than to shareholders 6.

In October 2025 the company completed a recapitalisation. The operating business became OpenAI Group PBC, a public benefit corporation; the non-profit, renamed the OpenAI Foundation, retained a stake of about 26 per cent and the right to appoint the board; Microsoft received about 27 per cent, reported to be worth roughly 135 billion dollars; the remainder went to employees and other investors 2.

The Microsoft relationship is both a funding arrangement and a dependency. Microsoft supplies Azure compute, resells OpenAI models, and has integrated them into its own products; OpenAI, in turn, has sought to diversify its suppliers and has signed large computing agreements with other providers. The two companies renegotiated their agreement during the 2025 restructuring, loosening some exclusivity while preserving Microsoft's stake and licences 2.

Private funding has been extraordinary even by the standards of the sector. A round in October 2024 raised 6.6 billion dollars at a valuation of 157 billion; a SoftBank-led round in March 2025 raised 40 billion at a reported valuation of 300 billion. The company says the capital is needed for compute, and its spending plans assume continued access to debt and equity markets.

OpenAI's founding documents commit it to safe development, and its public safety apparatus has evolved through several iterations. The Preparedness Framework, published at the end of 2023, defines capability thresholds in four areas — cyber, biological, persuasion and autonomy — and requires evaluation before deployment and, at the highest level, board review 1.

The company also runs a red-team programme, publishes system cards with each major model, and pays bounties for jailbreaks. It has argued publicly for licensing and testing requirements for frontier models, a position that places it closer to regulators than some of its competitors 1.

The framework has been revised since, and the company has added external review for the highest capability tier. It published a Model Spec in 2024 describing the behaviour it wants from its models, and it has allowed government evaluators in the United States and Britain to test some systems before release 1.

The internal record is more contested. The superalignment team, created in 2023 to work on controlling systems more capable than their supervisors, was dissolved in 2024 after the departure of several senior researchers, including Jan Leike and Ilya Sutskever. Leike said publicly that safety culture had taken second place to product; the company disputed that characterisation and said the work had been distributed to other teams 6.

The board crisis of November 2023 is the defining internal dispute. On 17 November the board removed Sam Altman as chief executive, saying he had not been consistently candid in his communications; Greg Brockman resigned in response. After staff threatened to leave and investors pressed the board, Altman returned on 21 November under a new board chaired by Bret Taylor, and most of the previous directors departed 5.

The copyright litigation is the most consequential external one. On 27 December 2023 The New York Times sued OpenAI and Microsoft, alleging that the companies copied millions of articles to train models that now compete with the newspaper, and that outputs could reproduce substantial passages. OpenAI has argued that training is fair use and that the passages reproduced in the complaint resulted from an atypical prompting method 4.

Regulators moved earlier. Italy's data protection authority temporarily blocked ChatGPT on 31 March 2023, citing the absence of a lawful basis for processing personal data and inadequate age verification. The service returned in April after OpenAI added disclosures and controls, and the authority later imposed a fine. Other European regulators opened similar inquiries, and several governments restricted the service on official devices 7.

Smaller disputes have concerned content moderation, the treatment of voice actors and authors whose work appears in training data, and the company's lobbying. None has changed its direction, but together they have made OpenAI the most litigated and most closely regulated of the frontier laboratories.

The central disagreement concerns the 2025 restructuring. Supporters argue that a non-profit could not have funded frontier training, and that the foundation's stake, board appointment power and mission charter preserve public benefit. Critics, including some former employees, argue that a 26 per cent stake is too small to constrain a company valued in the hundreds of billions, and that the foundation has become a minority shareholder in all but name 2.

A second dispute concerns capability claims. OpenAI presents each release as a step toward general-purpose systems; independent evaluators have noted that the gaps between leading models have narrowed on most benchmarks, that contamination of test sets is widespread, and that vendor-reported results are not directly comparable across laboratories 3.

A third concerns safety. The company's framework permits deployment of high-capability models with mitigations rather than prohibition, which some researchers regard as reasonable engineering and others as a conflict of interest, since the organisation that builds a system also decides whether it is safe enough to sell 6.

These are disagreements about interpretation as much as about fact, and the record does not settle them. What is not disputed is that OpenAI's choices — about structure, deployment and litigation posture — have become the template against which other laboratories are judged.

OpenAI remains the reference point for consumer AI. ChatGPT has the largest user base of any general assistant, its API is the default integration for a large share of start-ups, and its releases set the terms of comparison for competitors in the United States, Europe and China 1.

The position is not secure. Anthropic has taken a large share of enterprise spending, Google reaches users through search, Android and Workspace, and open-weight models from Meta, DeepSeek and Alibaba have compressed the price of capable inference. The company's response has been to widen its product surface and to spend heavily on compute and talent.

Its commercial base is lopsided. A small number of very large customers and a very large number of small developers account for much of the revenue, a distribution that makes the business sensitive both to a single customer's decisions and to price competition from models that can be downloaded and run without a vendor.

For an encyclopedia, OpenAI's significance is double. It commercialised a technology that had been a research programme, and it did so through an ownership structure designed to resist commercial pressure — a structure that it then rebuilt when the pressure proved stronger than expected 2.