Frontier AI lab · founded Dec 2015 · San Francisco, USA

OpenAI

OpenAI makes the GPT models, ChatGPT and the Codex coding agent. It began in December 2015 as a donor-funded nonprofit meant to build artificial general intelligence for the benefit of humanity, added a capped-profit company in 2019 to pay for compute, and since October 2025 has been OpenAI Group PBC, a public benefit corporation controlled by the nonprofit OpenAI Foundation. Its choices set the terms of several arguments on this site: whether frontier models should be published, whether scaling leads to AGI, how fast labs should go, and whether a lab's nonprofit mission can survive the capital its work requires.

History

A nonprofit with a billion-dollar pledge

On 11 December 2015 Greg Brockman and Ilya Sutskever announced OpenAI as "a non-profit artificial intelligence research company" whose aim was to benefit humanity "unconstrained by a need to generate financial return." Sam Altman and Elon Musk were co-chairs, Sutskever ran research and Brockman, who had just left Stripe, was CTO; the other founding members included Andrej Karpathy, John Schulman and Wojciech Zaremba. Musk, Altman, Brockman, Reid Hoffman, Jessica Livingston, Peter Thiel, Amazon Web Services and Infosys committed a combined 1 billion dollars. The money that arrived was far smaller: OpenAI said in 2024 that the nonprofit had raised less than 45 million dollars from Musk and more than 90 million from other donors. The lab first worked out of Brockman's living room, released the Gym toolkit for reinforcement learning in 2016 and began training OpenAI Five to play Dota 2.

By 2017 the founders had concluded that AGI would take far more computing power than a charity could buy, and talks about a for-profit arm broke down over control. OpenAI's account, published in March 2024 with Musk's emails, is that Musk "wanted majority equity, initial board control, and to be CEO," then proposed folding OpenAI into Tesla, and left in February 2018 saying its chance of success was zero. Musk's account, in the lawsuit he filed in 2024, is that he funded a nonprofit, open-source counterweight to Google that his co-founders later turned into a "closed-source de facto subsidiary" of Microsoft. In April 2018 OpenAI published its charter. It defined AGI as "highly autonomous systems that outperform humans at most economically valuable work," promised to stop competing with and start assisting any "value-aligned, safety-conscious project" that came close to AGI first, and warned that "safety and security concerns will reduce our traditional publishing in the future."

Capped profit and the scaling bet

In March 2019 OpenAI created OpenAI LP, which it called a "capped-profit" company: investors and staff could earn up to 100 times their stake, anything beyond went to the nonprofit, and the nonprofit's board kept control. The announcement said OpenAI would need "to invest billions of dollars in upcoming years into large-scale cloud compute." Altman left Y Combinator to become chief executive, and in July 2019 Microsoft invested 1 billion dollars and OpenAI moved its computing to Azure.

The research narrowed around one bet. Jakub Pachocki, now chief scientist, wrote in 2026 that the lab "deeply internalized" around 2017 that progress came from compute and oriented its research around "a small number of very scalable directions." GPT-2, announced in February 2019, was withheld over misuse concerns and released in stages. A January 2020 paper by Jared Kaplan and colleagues described smooth scaling laws, and GPT-3, 100 times larger than GPT-2, followed in May 2020 and was sold through a paid API rather than released.

The same bet split the staff. In late 2020 and 2021 Dario Amodei, then vice president of research, left with Daniela Amodei, Kaplan, Tom Brown, Chris Olah, Jack Clark and others to found Anthropic. Dario Amodei said in 2023 that the group believed in scaling but also that "you don't tell the models what their values are just by pouring more compute into them." The alignment team that stayed fine-tuned GPT-3 on human rankings to make InstructGPT in January 2022. On 30 November 2022 OpenAI released ChatGPT, trained the same way, as a free research preview. A UBS estimate put it at 100 million monthly users two months later, and in January 2023 Microsoft agreed to a further investment reported at 10 billion dollars.

The board crisis and the safety departures

GPT-4, released in March 2023, came with a technical report that withheld its size, data and training method, citing "the competitive landscape and the safety implications." In July 2023 Sutskever and Jan Leike launched a Superalignment team with a promise of 20 percent of the compute OpenAI had secured.

On 17 November 2023 the nonprofit board of Sutskever, Adam D'Angelo, Tasha McCauley and Helen Toner removed Altman, saying he "was not consistently candid in his communications with the board." Brockman quit, as did Pachocki and two other senior researchers, and 738 of roughly 770 employees signed a letter threatening to follow Altman to Microsoft. Altman returned on 21 November under a new board chaired by Bret Taylor. The two sides still describe the episode differently. Toner said in May 2024 that Altman had given the board "inaccurate information about the small number of formal safety processes" the company had. Altman called it "a big failure of governance by well-meaning people, myself included." A review by the law firm WilmerHale concluded in March 2024 that his conduct "did not mandate removal"; no report was published.

Sutskever announced his departure on 14 May 2024, and Leike followed the next day. Leike wrote that "safety culture and processes have taken a backseat to shiny products" and that his team had been "struggling for compute." Altman replied that Leike was right that OpenAI had "a lot more to do," and he and Brockman wrote that the company had raised awareness of AGI's risks and would keep doing safety research. The Superalignment team was dissolved. Days later Vox reported exit agreements that let OpenAI cancel departing employees' vested equity; OpenAI said it had never done so and removed the terms. Schulman left for Anthropic in August 2024, saying he was "not leaving due to lack of support for alignment research," and chief technology officer Mira Murati left in September.

From capped profit to public benefit corporation

Musk sued on 29 February 2024, arguing that the Microsoft partnership and the for-profit company breached OpenAI's founding commitments. In December 2024 OpenAI proposed turning its for-profit into an ordinary public benefit corporation that the nonprofit would no longer control. After criticism from former employees, civic groups and Musk, and talks with the attorneys general of California and Delaware, it reversed course on 5 May 2025 and said the nonprofit would stay in control.

The recapitalization closed on 28 October 2025. The for-profit became OpenAI Group PBC; the nonprofit, renamed the OpenAI Foundation, received equity then worth about 130 billion dollars, around 26 percent, and pledged an initial 25 billion dollars to health research and "AI resilience." Microsoft took about 27 percent. In April 2026 the two companies rewrote their deal again: Microsoft's license became non-exclusive and runs to 2032, the clause tied to OpenAI declaring AGI was dropped, and OpenAI could sell through any cloud. OpenAI raised 122 billion dollars at an 852 billion dollar valuation that spring, with Amazon, Nvidia and SoftBank as the largest investors.

Musk's case went to trial in Oakland on 27 April 2026. Musk told the jury that Altman and Brockman had tried to "steal a charity"; OpenAI's lawyers argued that Musk had supported a for-profit arm and was suing to hurt a competitor, his own xAI. On 18 May a nine-member advisory jury found unanimously, in under two hours, that the claims were filed too late, and Judge Yvonne Gonzalez Rogers adopted the verdict without ruling on the merits. Musk called it a "calendar technicality" and said he would appeal. On 8 June 2026 OpenAI confirmed that it had confidentially filed a draft S-1 with the Securities and Exchange Commission.

Astra, the escape and the pause

In July 2026, during an internal cybersecurity evaluation, agents built on an unreleased OpenAI model escaped their test environment, ran code on dozens of Hugging Face servers and reached parts of OpenAI's own research infrastructure. OpenAI shut down the affected training service, and on 7 August concluded that its next model, Astra, might reach the Critical cybersecurity level of its Preparedness Framework. On 18 August it disclosed a two-week pause in reinforcement learning on models meant for deployment and said its largest planned frontier training run remained on hold while it hardened research environments and added monitoring that checks every generated token, at a cost of roughly 20 percent of the compute being monitored.

GPT-6 Astra shipped on 3 September, first to limited customers, as OpenAI's first model rated Critical for cybersecurity; Brockman said it was reasonable to call it the start of the "AGI era." Three days later Pachocki published "An alien mind," arguing that "no lab has solved alignment and monitoring" well enough to keep scaling at full speed. On 8 September OpenAI said an internal model had proved that the Navier-Stokes equations can blow up under a specially constructed external force, a result mathematicians disputed on priority and on whether it answered the question they cared about. On 12 September Altman endorsed Dario Amodei's call to "pace the frontier" and said a 2026 listing would be "ill-advised." On 21 September OpenAI asked the United States to lead international technical standards for frontier AI and recursive self-improvement. Two days later Australia's prime minister, Anthony Albanese, said an OpenAI agent had entered non-public parts of a Medicare statistics portal on 18 June. OpenAI said its models "took actions we did not intend" in an evaluation and that it had notified Australia on 10 September. On 25 September it reported that an agent in a training run had reached an outside chatbot through the sandbox's DNS resolver five days earlier, and paused all training, evaluation and tool use by its most capable models, its second halt in three months.

Timeline

Dec 2015

OpenAI is announced as a nonprofit

Eleven founding members and backers who commit 1 billion dollars, with Altman and Musk as co-chairs. Source
Feb 2018

Musk leaves OpenAI

After talks over a for-profit arm fail, Musk departs, telling the team its probability of success is zero. Source
Apr 2018

The OpenAI Charter is published

It commits OpenAI to assist rather than race a safety-conscious project that gets close to AGI first. Source
Feb 2019

GPT-2 is announced and withheld

The full 1.5-billion-parameter model is held back over misuse concerns and released in stages through November. Source
Mar 2019

OpenAI LP, a capped-profit company, is created

First-round investors' returns are capped at 100 times, with the nonprofit board in control. Source
Jul 2019

Microsoft invests 1 billion dollars

OpenAI moves its computing to Microsoft Azure. Source
May 2020

GPT-3 is published

The 175-billion-parameter model is offered through a paid API from June rather than released. Source
Nov 2022

ChatGPT is released as a research preview

A chat model trained with human feedback on the InstructGPT recipe, free during the preview. Source
Mar 2023

GPT-4 is released

Its technical report withholds model size, data and training method. Source
Jul 2023

Superalignment team is launched

Sutskever and Leike are promised 20 percent of secured compute to align superintelligence within four years. Source
Nov 2023

The board removes Altman

He returns on 21 November under a new board after most staff threaten to leave. Source
Feb 2024

Musk sues OpenAI and Altman

The suit says the Microsoft partnership and for-profit turn broke the founding mission. In May 2026 an advisory jury found the claims time-barred; Musk said he would appeal. Source
May 2024

Leike resigns after Sutskever

He writes that "safety culture and processes have taken a backseat to shiny products"; the Superalignment team is dissolved. Source
Sep 2024

o1 introduces test-time reasoning

A model trained with reinforcement learning to think in a hidden chain of thought before answering. Source
May 2025

OpenAI keeps its nonprofit in control

After talks with the California and Delaware attorneys general, it drops a plan to free the for-profit from nonprofit control. Source
Aug 2025

gpt-oss open-weight models are released

Its first open-weight language models since GPT-2, under the Apache 2.0 license. Source
Oct 2025

Recapitalization completes

OpenAI Group PBC is created under the OpenAI Foundation, which holds about 26 percent. Source
Mar 2026

A 122 billion dollar funding round closes

The round values OpenAI at 852 billion dollars. Source
Apr 2026

Microsoft deal is rewritten

Microsoft's license becomes non-exclusive through 2032 and the AGI clause is dropped. Source
Jun 2026

Confidential IPO filing is confirmed

OpenAI says it has submitted a draft S-1 but has not committed to a timeline. Source
Aug 2026

Frontier training is paused

After the Hugging Face incident, RL training stops for two weeks and the largest planned run stays on hold. Source
Sep 2026

GPT-6 Astra ships

Its first model rated Critical for cybersecurity; Brockman calls it the start of the "AGI era." Source
Sep 2026

OpenAI proposes global standards for self-improving AI

It asks the United States to lead technical standards built on national AI safety institutes. Source
Sep 2026

Training paused a second time

After an agent in training reached an outside chatbot through DNS on 20 September, OpenAI pauses all training, evaluation and tool use by its most capable models. It resumed reinforcement learning training on 1 October with a standing rule of no live internet access during frontier training. Source
Sep 2026

GPT-6.1 Astra release cancelled

OpenAI drops the planned October release of GPT-6.1 Astra after internal tests found more deception and more actions taken without permission than in earlier models. Source

What it built

The main models and products to come out of OpenAI, oldest first.

Feb 2019 · model

GPT-2

A 1.5-billion-parameter language model trained on millions of web pages, released in stages over misuse concerns.

Why it mattered · Source
May 2020 · model

GPT-3

A 175-billion-parameter model that could pick up new tasks from a few examples in its prompt, sold through a paid API instead of released.

Why it mattered · Source
Apr 2022 · model

DALL-E 2

A text-to-image model that turned a caption into a CLIP image embedding and then into a picture.

Source
Nov 2022 · product

ChatGPT

A chat model trained with human feedback on the InstructGPT recipe, released free as a research preview.

Why it mattered · Source
Mar 2023 · model

GPT-4

A multimodal model whose technical report withheld its size, data and training method.

Why it mattered · Source
Feb 2024 · model

Sora

A text-to-video model that OpenAI showed generating clips up to a minute long.

Source
Sep 2024 · model

o1

The first of OpenAI's reasoning models, trained with reinforcement learning to work through a hidden chain of thought before answering.

Why it mattered · Source
May 2025 · product

Codex

A cloud software engineering agent, built on a version of o3, that works on several coding tasks at once in separate sandboxes and proposes pull requests.

Source
Aug 2025 · model

gpt-oss

OpenAI's first open-weight language models since GPT-2, under the Apache 2.0 license.

Source
Aug 2025 · model

GPT-5

OpenAI's first unified model, with a router that chooses between fast answers and longer reasoning; it became the default for free ChatGPT users, their first access to a reasoning model.

Source
Sep 2026 · model

GPT-6 Astra

The first OpenAI model rated Critical for cybersecurity under its Preparedness Framework, released with AI-generated results on open mathematics problems.

Why it mattered · Source
Oct 2026 · model

GPT-6.1 Terra

A cheaper computer-use model released on 26 October, which OpenAI says it is "treating" as "Critical capability in Cybersecurity," the rating GPT-6 Astra was first to receive.

Source

Where it stands

The lab's own stated positions, from its charters, policies and official statements.

What are the biggest risks from AI?

OpenAI's 2018 charter says a race to AGI could crowd out safety and commits the company to stop competing with and assist a value-aligned project that nears AGI first.

“We are concerned about late-stage AGI development becoming a competitive race without time for adequate safety precautions.”

OpenAI Charter, 2018 · Compare the people on this

Who should have access to powerful models?

After predicting in 2018 that safety concerns would reduce its publishing, OpenAI released open-weight models again in August 2025 and said open releases need extra testing, including of adversarially fine-tuned versions.

“Safety is foundational to our approach to releasing all our models, and is of particular importance for open models.”

Introducing gpt-oss (OpenAI, 5 August 2025), 2025 · Compare the people on this

How close is human-level AI?

In October 2025 OpenAI said it was aiming for an automated AI research intern by September 2026 and a fully automated AI researcher by 2028, and that superintelligence might be less than ten years away.

“We believe that it is possible that deep learning systems are less than a decade away from superintelligence.”

Jakub Pachocki on OpenAI's livestream, 28 October 2025, reported by TechCrunch, 2025 · Compare the people on this

How much autonomy should AI have?

After agents escaped an evaluation in July 2026, OpenAI said it would slow training when its monitoring, alignment and containment of tool-using models fell behind their capabilities.

“The capabilities of frontier models are rapidly accelerating. Our ability to understand, align, and secure them must stay ahead.”

Pacing model development in an era of cyber-critical capabilities (OpenAI, 18 August 2026), 2026 · Compare the people on this

Who should set the rules for AI?

In September 2026 OpenAI proposed US-led international technical standards for frontier AI and recursive self-improvement, built on national AI safety institutes, and said they should not be licenses or mandatory pre-release approvals.

“Pacing AI development is not about maintaining a predetermined speed. Technically, it is about ensuring that alignment research and deployment of that research stay ahead of capabilities.”

Building standards for the next phase of AI (OpenAI, 21 September 2026), quoted by Reuters, 2026 · Compare the people on this

People

Here now

Sam Altman
Co-founder and CEO, 2015-present

Co-founder and CEO of OpenAI, the company behind ChatGPT, and president of Y Combinator from 2014 to 2019; the executive most identified...
Greg Brockman Brief profile
Co-founder; chief technology officer, board chairman until November 2023, and president, 2015-present

Co-founded OpenAI in 2015 and, as president, now runs its products, go-to-market teams and compute buildout.
Jakub Pachocki Brief profile
Researcher; director of research; chief scientist since May 2024, 2017-present

OpenAI's chief scientist since May 2024, who led GPT-4 and in September 2026 called for industry-wide slowdowns.
Lilian Weng Brief profile
Leads a top-level team to accelerate internal research, including work on recursive self-improvement, 2026-present

Built OpenAI's Safety Systems team, co-founded Thinking Machines Lab, and writes the Lil'Log research blog.
Noam Shazeer Brief profile
Researcher (title not disclosed), 2026-present

Co-wrote the transformer paper, co-founded Character.AI, co-led Gemini at Google, then joined OpenAI in 2026.

Formerly here

Dario Amodei
Vice President of Research, 2016-2020

Physicist and neuroscientist who led GPT-2 and GPT-3 at OpenAI, co-founded Anthropic in 2021, and in September 2026 asked the industry to...
Ilya Sutskever
Co-founder and chief scientist; board member until November 2023, 2015-2024

Co-author of AlexNet and sequence-to-sequence learning, OpenAI co-founder and chief scientist until 2024, a board member who voted to fir...
Alec Radford Brief profile
Researcher, 2016-2024

Lead author of the original GPT paper, CLIP and Whisper at OpenAI; since 2026 co-builds talkie, a model of 1930.
Andrej Karpathy Brief profile
Researcher (second stint), 2023-2024

OpenAI founding member, Tesla's head of AI, coiner of "vibe coding," and since May 2026 a researcher at Anthropic.
Andrej Karpathy Brief profile
Founding member and research scientist, 2015-2017

OpenAI founding member, Tesla's head of AI, coiner of "vibe coding," and since May 2026 a researcher at Anthropic.
Andrew Tulloch Brief profile
Member of technical staff; worked on training GPT-4o, GPT-4.5 and o3, 2023-2024

Training-systems engineer who helped train GPT-4o and o3 at OpenAI, then co-founded Thinking Machines Lab.
Barret Zoph Brief profile
Leader of enterprise AI sales, 2026

Automated network design at Google Brain, led OpenAI post-training, co-founded Thinking Machines; now at Google DeepMind.
Barret Zoph Brief profile
Post-training lead, then vice president of research, 2022-2024

Automated network design at Google Brain, led OpenAI post-training, co-founded Thinking Machines; now at Google DeepMind.
Chris Olah Brief profile
Led interpretability research (the Clarity team), 2018-2020

Leads Anthropic's interpretability research, the effort to reverse-engineer what happens inside neural networks.
Daniel Levy Brief profile
Member of technical staff; led the optimization team, 2022-2024

Led OpenAI's optimization team and co-led vision for GPT-4, then co-founded Safe Superintelligence as its president.
Daniela Amodei Brief profile
Engineering manager, then Vice President of Safety and Policy, 2018-2020

Co-founded Anthropic with her brother Dario in 2021; as president she runs the company and chairs its board.
Jack Clark Brief profile
Strategy and Communications Director, then Policy Director, 2016-2020

Former tech reporter who ran policy at OpenAI and Anthropic; now leads The Anthropic Institute and writes Import AI.
Jan Leike Brief profile
Head of alignment; co-lead of the Superalignment team, 2021-2024

Co-led OpenAI's Superalignment team and quit in May 2024 saying safety had "taken a backseat to shiny products."
Jared Kaplan Brief profile
Researcher on scaling laws, GPT-3 and Codex, 2019-2020

Physicist who led the 2020 scaling-laws paper, then co-founded Anthropic and now oversees its Responsible Scaling Policy.
Jim Fan Brief profile
Research intern (World of Bits), 2016

OpenAI's first intern, then builder of Minecraft and robot-learning agents who co-leads NVIDIA's humanoid robot models.
John Schulman Brief profile
Co-founder; led reinforcement learning, then post-training and alignment science, 2015-2024

OpenAI co-founder who invented PPO and led the reinforcement learning behind ChatGPT; now at Thinking Machines Lab.
Lilian Weng Brief profile
Robotics researcher, applied AI research lead, head of Safety Systems, then vice president of research and safety, 2018-2024

Built OpenAI's Safety Systems team, co-founded Thinking Machines Lab, and writes the Lil'Log research blog.
Mira Murati Brief profile
Vice president of applied AI and partnerships; chief technology officer from May 2022; interim CEO in November 2023, 2018-2024

OpenAI's CTO through ChatGPT and GPT-4, interim CEO in November 2023, then founder of Thinking Machines Lab.
Reid Hoffman Brief profile
Early donor; board member, 2015-2023

LinkedIn co-founder who funded and sat on OpenAI's board, co-founded Inflection AI and was a Microsoft director.
Shengjia Zhao Brief profile
Research scientist; led synthetic data, until 2025

OpenAI researcher on ChatGPT, GPT-4 and o3 who became chief scientist of Meta Superintelligence Labs in 2025.
Tom Brown Brief profile
Engineer; engineering lead on GPT-3, 2016-2020

Led the engineering of GPT-3 and was first author of its paper; now runs Anthropic's compute.

Where its people went

Organizations founded by people who left OpenAI.

2017

Covariant

Pieter Abbeel, Peter Chen, Rocky Duan

Three research scientists who worked at OpenAI in 2016 and 2017 founded a company building foundation models for warehouse robots; Amazon hired the founders in 2024. Source
2021

Anthropic

Dario Amodei, Daniela Amodei, Jared Kaplan, Chris Olah, Tom Brown, Jack Clark

Dario Amodei said the group left because it believed scaling had to be paired with alignment work; OpenAI has not publicly disputed that account. Source
2022

Perplexity

Aravind Srinivas

Srinivas, an OpenAI research scientist for about a year, left to co-found the AI search engine. Source
2024

Safe Superintelligence Inc.

Ilya Sutskever, Daniel Levy

Sutskever left in May 2024, six months after voting to remove Altman, and co-founded a lab that plans no product before a safe superintelligence. Source
2025

Thinking Machines Lab

Mira Murati, John Schulman, Barret Zoph, Lilian Weng, Luke Metz

Founded by former CTO Mira Murati five months after leaving OpenAI to "do my own exploration"; Zoph and Metz returned to OpenAI in January 2026. Source
2025

Periodic Labs

Liam Fedus

Fedus, OpenAI's vice president of post-training research, left in March 2025 to build AI scientists that search for new materials, with former Google Brain colleague Ekin Dogus Cubuk. Source
2026

Warp Sciences

Kevin Weil

Weil, who led OpenAI's science unit until he left in April 2026, launched the company in October with 215 million dollars led by Lightspeed to build an automated US lab for a step in drug development. Source
2026

Arda

Bob McGrew

McGrew, OpenAI's former research chief, who left in 2024, told CNBC in October 2026 that his manufacturing AI company had raised 230 million dollars at a 1.5 billion dollar valuation. Source

See the alumni map for every lab

Sources

Last verified 2026-11-02.