OpenAI

OpenAI is the artificial-intelligence research and deployment organization behind ChatGPT, the OpenAI API, and Codex. Its work connects large-scale model training with conversation, tool use, and delegated software tasks. Its evolution includes consequential changes in financing and governance, alongside advances in specifying tasks through examples, training models to follow instructions, and spending additional computation on reasoning.[10][7][14][6][9][12][3][2]

Following the restructuring announced on October 28, 2025, OpenAI comprises the nonprofit OpenAI Foundation and the for-profit OpenAI Group PBC, a public benefit corporation. The Foundation controls the Group through special voting and governance rights, including the power to appoint and replace all Group directors. As of September 17, 2026, its official structure page identifies Sam Altman as CEO and Bret Taylor as Foundation chair.[2]

This arrangement followed a nonprofit-controlled, capped-profit partnership established in 2019. OpenAI’s earlier founding commitments and shared research infrastructure supply context for that financing change and the commercial API that followed.[1][4][3][7]

OpenAI introduced itself publicly on December 11, 2015, as a nonprofit research company pursuing broadly beneficial AI without a need to generate financial returns. It encouraged researchers to publish papers, blog posts, and code and committed to collaboration across institutions. Founding-era leadership included Ilya Sutskever as research director, Greg Brockman as CTO, and Sam Altman and Elon Musk as co-chairs. Other founding members included Trevor Blackwell, Vicki Cheung, Andrej Karpathy, Durk Kingma, John Schulman, Pamela Vagata, and Wojciech Zaremba. The announcement described $1 billion in aggregate funder commitments.[1]

OpenAI Gym, released in public beta on April 27, 2016, began as infrastructure for OpenAI’s own reinforcement-learning research. It supplied reusable environments, including Atari games and simulated robots, for developing and comparing algorithms across machine-learning frameworks. Versioned environments preserved consistent task definitions: changing an environment created a new version, helping researchers reproduce and compare results.[4]

On March 11, 2019, OpenAI said it had “decided to scale much faster than we’d planned when starting OpenAI.” It created a nonprofit-controlled, capped-profit partnership to finance compute, talent, and supercomputers. First-round investor returns were capped at 100 times their investment, with excess returns belonging to the nonprofit. At the time, OpenAI reported approximately 100 employees and emphasized technology research rather than commercial products. Musk had left the nonprofit board in February 2018 and was not formally involved in OpenAI LP.[3]

On July 22, 2019, OpenAI announced a $1 billion Microsoft investment, joint development of Azure AI supercomputing infrastructure, and Microsoft as its exclusive cloud provider. Its commercial API followed in private beta on June 11, 2020, serving GPT-3-family models. OpenAI described paid deployment as a way to finance research, make expensive models accessible without local operation, learn from real applications, and adjust access in response to misuse.[17][7]

GPT-2, announced in February 2019, was a 1.5-billion-parameter Transformer trained for next-word prediction. OpenAI initially withheld the largest model because of potential misuse, treating disclosure as an experiment and acknowledging uncertainty: “we are not sure that it is the right decision today.” After staged releases and partner sharing, it released the largest model on November 5, 2019.[5][18]

The 2020 GPT-3 paper described a 175-billion-parameter autoregressive language model. Tasks were specified through instructions and demonstrations in the input, without task-specific gradient updates during evaluation. A developer could provide examples of a text transformation and request another; the initial API exposed this through a “text in, text out” interface. In-context task specification enabled adaptation through prompts, although results varied across tasks. The researchers also distinguished recognizing previously learned patterns from learning a new task from scratch at inference time.[6][7]

OpenAI introduced CLIP and DALL·E on January 5, 2021. CLIP used contrastive training to associate images with paired text, allowing candidate natural-language class descriptions to specify classification tasks without task-specific training examples. It built on earlier visual-language research, including work by Richard Socher and Stanford collaborators and Ang Li and FAIR collaborators. It remained sensitive to label wording, with weaknesses in counting and fine-grained categories. Original DALL·E instead generated images by modeling text and image tokens autoregressively in one sequence. Its interactive visuals displayed the top 32 of 512 candidates after CLIP reranking.[19][20]

DALL·E 2 extended image generation with inpainting, outpainting, and variations. OpenAI moved it from a limited trusted-user preview to beta availability in July 2022. Research advancements were credited to Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen, alongside separately credited engineering, product, policy, and safety contributors.[21]

On September 21, 2022, OpenAI released Whisper’s models and inference code as open source. Trained on 680,000 hours of multilingual and multitask supervised audio, Whisper supported transcription and translation into English. Its encoder-decoder Transformer processed 30-second audio chunks as log-Mel spectrograms, with decoder tokens directing language identification, phrase-level timestamps, transcription, and translation.[22]

The 2022 InstructGPT research, a joint OpenAI Alignment-team project by Long Ouyang, Jeff Wu, and collaborators, addressed the difference between predicting a continuation and following a request. Human demonstrations supported supervised fine-tuning; ranked model outputs trained a reward model, which then guided reinforcement learning. On the authors’ prompt distribution, human evaluators preferred a 1.3-billion-parameter InstructGPT model to 175-billion-parameter GPT-3. The result showed the importance of feedback within that evaluation. The authors explicitly bounded alignment to their labelers’ and researchers’ preferences rather than a broader notion of human values.[9]

ChatGPT launched on November 30, 2022, as a free research preview fine-tuned from GPT-3.5. Human trainers played both sides of dialogues; rankings of alternative responses trained reward models for subsequent reinforcement learning. The conversational interface supported follow-up questions and refined requests. OpenAI also warned that ChatGPT could produce plausible but incorrect answers, respond differently to prompt wording, and guess at ambiguous intent.[10]

GPT-4, announced on March 14, 2023, accepted image and text inputs and produced text. Text capability launched through ChatGPT Plus and API access with a waitlist; image input was initially restricted, with Be My Eyes named as the initial partner. OpenAI described next-word pretraining followed by human-feedback post-training, using a supercomputer co-designed with Azure. It also released OpenAI Evals for evaluation and regression detection; Stripe used it alongside human evaluation of a GPT-powered documentation tool. The launch acknowledged hallucinations, reasoning errors, biases, and imperfect safeguards.[23]

The June 13, 2023 function-calling release connected model outputs to external tools. Developers described functions through JSON Schema, and models selected arguments in a JSON object. In a weather application, the model could select a location while application code performed the lookup. Generating arguments and executing an action are separate steps; the announcement warned that untrusted tool outputs could direct unintended actions.[11]

The release reduced embeddings pricing by 75% and GPT-3.5 input-token pricing separately by 25%. A July 20 update extended support for older pinned models after customer feedback. OpenAI acknowledged that aggregate improvements could coexist with regressions on particular tasks, giving developers a reason to evaluate model changes against their own applications.[11]

On November 17, 2023, OpenAI’s board removed Altman as CEO and director, stating that “he was not consistently candid in his communications with the board.” It appointed Mira Murati interim CEO. OpenAI’s November 29 announcement confirmed Altman’s return, with an initial board of Bret Taylor, Larry Summers, and Adam D’Angelo. Murati returned as CTO and Brockman as president. The later governance review dated the rehiring decision to November 21.[25][26][27]

In March 2024, OpenAI published a summary of WilmerHale’s commissioned review. According to that summary, WilmerHale attributed the removal to a breakdown of trust between the prior board and Altman, rather than concerns about product safety, development pace, finances, or statements to outside stakeholders. It found that the board had broad discretion to terminate him, but that his conduct did not mandate removal. OpenAI endorsed Altman and Brockman’s continuing leadership and announced governance guidelines, a strengthened conflicts policy, a whistleblower hotline, and additional board committees.[27]

OpenAI’s February 15, 2024 Sora report described a text-conditional diffusion Transformer operating on spacetime patches of compressed video and image representations. A compression network mapped video into latent representations; Sora learned to denoise patches in that space, with generated latents decoded back into pixels. This representation accommodated different durations, resolutions, and aspect ratios. OpenAI reported minute-long generation by its largest research model while withholding implementation details, and presented physical-world simulation as a research direction.[24]

The September 12, 2024 o1 report introduced additional computation spent reasoning before answering as another capability lever. OpenAI reported that reinforcement learning improved reasoning strategies and that performance increased with both training compute and inference-time compute. Its research model and released o1-preview were distinct; reported benchmark results depended on compute and sampling settings. OpenAI withheld raw reasoning traces after considering user experience, competitive advantage, and monitoring, presenting model-generated summaries instead.[12]

The March 11, 2025 Responses API and Agents SDK launch integrated model turns, web search, file search, computer use, orchestration, and workflow tracing. OpenAI described these tools as addressing customers’ difficulties with prompt iteration and custom orchestration. Repeated model and tool interactions could form a workflow, with tracing exposing its steps for inspection.[13]

Codex spans two distinct generations. The 2021 paper, by Mark Chen, Jerry Tworek, Heewoo Jun, and collaborators, described a GPT model fine-tuned on public GitHub code; a separate production version powered GitHub Copilot. Its HumanEval benchmark used 164 handwritten problems and unit tests to evaluate programs generated from docstrings. The May 16, 2025 Codex launch instead introduced a cloud agent that worked in separate repository-loaded sandboxes, edited files, ran commands and tests, and returned changes with terminal logs and test outputs for review. Its launch model, codex-1, was an o3 variant optimized for software engineering through reinforcement learning. A June update enabled users to grant internet access during tasks.[8][14]

By September 2026, Codex supported use across ChatGPT, editors, and the terminal, with worktrees, team-specific Skills, and scheduled background tasks. OpenAI’s GPT-6 Astra announcement described staged distribution and an experimental Codex feature for longer sessions: retaining notes across context windows and searching earlier messages and tool outputs for requirements or test results that the notes omitted.[15][16]

OpenAI introduced Frontier on February 5, 2026, as a platform for enterprises to build, deploy, and manage AI agents. In an April 7 Latent Space interview, Ryan Lopopolo identified his work as Frontier Product Exploration, developing new products in the space of OpenAI Frontier.[30][31]

Symphony, described in OpenAI’s April 27, 2026 announcement, is a written specification for supervising and orchestrating agentic work. Its repository offers both specification-based implementation and an experimental Elixir reference implementation under Apache License 2.0. As of September 18, the project was an engineering preview for testing in trusted environments.[35][34]

Symphony’s documented demonstration monitors a Linear board, dispatches work to coding agents, and returns CI status, PR review feedback, complexity analysis, and walkthrough videos. Accepted work is then landed as pull requests. The workflow shifts the interface from supervising each coding agent to assigning work and reviewing its results.[34]

The 2025 recapitalization replaced the historical capped-profit arrangement with conventional equity. At recapitalization closing, OpenAI recorded 26% ownership for the Foundation, roughly 27% for Microsoft, and the remaining 47% for current and former employees and investors. These closing-date shares preceded subsequent financing.[2][3]

On March 31, 2026, OpenAI announced that it had closed a funding round with $122 billion in committed capital at an $852 billion post-money valuation. It named Amazon, NVIDIA, and SoftBank as strategic anchors, with continued Microsoft participation. It also reported a cloud portfolio spanning Microsoft, Oracle, AWS, CoreWeave, and Google Cloud, expanding beyond the Microsoft exclusivity announced in 2019.[28][17]

In the same March announcement, OpenAI reported $2 billion in monthly revenue, more than 900 million ChatGPT weekly active users, over 50 million subscribers, and more than two million Codex weekly users. It said enterprise represented more than 40% of revenue and its APIs processed over 15 billion tokens per minute. These figures distinguish active use, paid subscriptions, revenue composition, and model throughput.[28]

By August 31, 2026, OpenAI reported more than one billion ChatGPT weekly active users. It also said ChatGPT Ads had reached $1 billion in annualized revenue run rate in fewer than 200 days, with tens of thousands of advertisers and availability in more than 40 countries. The advertising figure describes an annualized run rate, rather than a completed year’s revenue. OpenAI placed advertising alongside subscriptions, enterprise offerings, and usage-based APIs in its business model.[29]

In a September 30, 2026 interview after OpenAI DevDay, computer-use lead Ari Weinstein described agents operating applications through several forms of context. Instead of relying only on repeated screenshots and scrolling, a system can use accessibility information or a web page’s DOM, and can write code to perform several actions. He attributed improvements to both models and the surrounding software, including better recovery when an attempted workflow fails. He described these as OpenAI’s experience, not a universal reliability measurement.[36]

Weinstein said the Agents API exposed the computer-use implementation used in OpenAI products. He also described Dots as having access to individual cloud Linux computers, and pointed to App Shots as a way to supply both visual and accessibility context from an application. In his account, useful deployment requires controls suited to the task, including consent for consequential actions and limiting access to the applications an agent needs.[36]

In the same interview, API product lead Nikunj Handa described asynchronous tool calls: a model can start a slow tool, continue its work and receive results later. He also described mid-turn steering, where new instructions can arrive while a model is working, and bidirectional WebSockets as part of the supporting interface. These extend the agent loop beyond a sequence that must stop after every tool call.[36]

Handa said the team was building a Decisions API for fast classification after customer and internal interest in the pattern popularized by Jev. At the time of the interview, he described a prototype built on existing Luna weights, using constrained outputs, parallel questions and inference optimizations, with a launch planned after reaching a latency target. That statement describes the prototype and its planned release at the time, rather than proving its later availability or performance.[36]