CtrlK
| OpenAI | |
|---|---|
| * | |
| File:OpenAI Pioneer Building.jpg | |
| OpenAI headquarters in Pioneer Building, San Francisco | |
| Type | Public Benefit Corporation (OpenAI Group, formerly OpenAI Global LLC); Non-profit foundation (OpenAI Foundation) |
| Industry | Artificial intelligence |
| Founded | December 11, 2015 |
| Founders | Sam Altman Elon Musk Ilya Sutskever Greg Brockman Trevor Blackwell Vicki Cheung Andrej Karpathy Durk Kingma Jessica Livingston John Schulman Pamela Vagata Wojciech Zaremba |
| Headquarters | Pioneer Building, San Francisco, California, United States |
| Key people | Sam Altman (CEO) Greg Brockman (President) Sarah Friar (CFO) Brad Lightcap (COO) Jakub Pachocki (Chief Scientist) Kevin Weil (Chief Product Officer) Bret Taylor (Chairman) Jony Ive (Chief Design Officer) |
| Revenue | $13.1 billion (2025 actual); $25B+ annualized run rate (February 2026) |
| Valuation | $730 billion pre-money / $840 billion post-money (February 2026) |
| Employees | ~4,000+ (early 2026) |
| Website | openai.com |
See also: Organizations, Artificial Intelligence and Machine Learning*
OpenAI is an artificial intelligence (AI) research and deployment company founded on December 11, 2015, in San Francisco, California [1]. Originally structured as a non-profit with a capped-profit subsidiary, OpenAI completed a restructuring in October 2025, splitting into two entities: the OpenAI Foundation (a non-profit) and the OpenAI Group (a for-profit public benefit corporation) [2]. Its stated mission is to ensure that artificial general intelligence (AGI), highly autonomous systems that outperform humans at most economically valuable work, benefits all of humanity [3].
The company gained widespread public attention with the release of ChatGPT in November 2022, which reached 100 million users in just two months, becoming the fastest-growing consumer application in history [4]. Growth has continued at a remarkable pace: ChatGPT reached 400 million weekly active users by February 2025, 700 million by August 2025, and 900 million by February 2026 [5][6].
OpenAI has developed groundbreaking AI models including the GPT series of large language models, the o-series reasoning models, DALL-E image generation systems, Sora video generation technology, and Whisper speech recognition. In February 2026, the company closed a $110 billion funding round at a $730 billion pre-money valuation ($840 billion post-money), the largest private funding round in history, with investments from Amazon, Nvidia, and SoftBank [100]. An IPO is widely expected in late 2026 or early 2027 [8].
OpenAI was founded on December 11, 2015, as a non-profit artificial intelligence research laboratory by Sam Altman, Elon Musk, Ilya Sutskever, Greg Brockman, Trevor Blackwell, Vicki Cheung, Andrej Karpathy, Durk Kingma, Jessica Livingston, John Schulman, Pamela Vagata, and Wojciech Zaremba [1]. The founding members and early backers, including Reid Hoffman, Peter Thiel, Amazon Web Services, Infosys, and YC Research, collectively pledged over $1 billion to the venture, though only about $130 million was collected by 2019 [9].
The organization aimed to develop open-source AI research to mitigate existential risks from AGI and ensure its benefits would be widely distributed rather than concentrated in the hands of a few corporations [1]. Early projects included:
2016: Released OpenAI Gym, a toolkit for reinforcement learning research [10]
2016: Received a DGX-1 supercomputer from Nvidia, the first unit shipped
2016: Launched OpenAI Universe, a platform for training AI agents across various environments [11]
2018: Released GPT-1, introducing the transformer-based language model approach with 117 million parameters [12]
In February 2018, Elon Musk resigned from the board citing potential conflicts of interest with his work on AI at Tesla, though he remained a donor and advisor [13].
In March 2019, OpenAI underwent a significant structural transformation, creating OpenAI LP (later renamed OpenAI Global LLC), a "capped-profit" for-profit entity. This model limits investor returns to 100 times their investment, with excess profits flowing back to the non-profit parent [14]. This restructuring was motivated by the immense computational and capital requirements for developing AGI, estimated at billions of dollars.
In July 2019, Microsoft announced a $1 billion investment in OpenAI, marking the beginning of a strategic partnership [15]. The partnership provided:
Exclusive cloud computing through Microsoft Azure
Joint development of Azure AI supercomputing technologies
Microsoft as exclusive commercial partner for OpenAI technologies
Key releases during this period:
February 2019: GPT-2 with 1.5 billion parameters, initially withheld due to concerns about malicious use [16]
June 2020: GPT-3 with 175 billion parameters, demonstrating unprecedented few-shot learning capabilities [17]
September 2020: Microsoft licensed exclusive use of GPT-3 technology [18]
January 2021: DALL-E, first text-to-image generation model
April 2022: DALL-E 2 with improved resolution and capabilities
On November 30, 2022, OpenAI released ChatGPT, a conversational AI chatbot based on GPT-3.5 using Reinforcement Learning from Human Feedback (RLHF) [19]. The release represented a turning point in AI history:
Reached 1 million users in 5 days
100 million users in 2 months (fastest consumer app growth ever)
Sparked global conversation about AI capabilities and risks
Led to rapid AI development across the industry
In January 2023, Microsoft announced an expanded multi-year, multi-billion dollar investment (reported as $10 billion), deepening the partnership and integrating OpenAI models across Microsoft products including Bing and Microsoft 365 [20].
March 14, 2023, saw the release of GPT-4, a multimodal model capable of processing both text and images, showing human-level performance on various professional and academic benchmarks [21].
On November 17, 2023, OpenAI experienced an unprecedented governance crisis when the board of directors abruptly removed Sam Altman as CEO, citing that he "was not consistently candid in his communications with the board, hindering its ability to exercise its responsibilities" [22]. The board action was reportedly led by chief scientist Ilya Sutskever and board members Helen Toner, Tasha McCauley, and Adam D'Angelo [23].
The decision triggered immediate backlash:
President Greg Brockman resigned in protest
Over 700 of OpenAI's ~770 employees (>95%) signed an open letter threatening to resign and join Microsoft unless Altman was reinstated [24]
Microsoft CEO Satya Nadella announced Altman and Brockman would lead a new AI research unit at Microsoft
Major investors including Microsoft, Thrive Capital, and Sequoia Capital pressured for Altman's return
After five days of intense negotiations, on November 22, 2023, Altman was reinstated as CEO with a new initial board comprising:
Bret Taylor (Chairman, former co-CEO of Salesforce)
Larry Summers (former U.S. Treasury Secretary)
Adam D'Angelo (CEO of Quora, only remaining member from previous board) [25]
The crisis reportedly stemmed from fundamental disagreements about AI safety, commercialization pace, and the balance between the non-profit mission and for-profit operations [26]. An internal investigation concluded in March 2024 that Altman's behavior "did not mandate removal," and the company adopted new corporate governance guidelines, including a whistleblower hotline. Altman was returned to the board at that time, along with three new directors: Nicole Seligman, Fidji Simo, and Sue Desmond-Hellmann [27].
2024 was a year of rapid product launches, significant personnel changes, and historic fundraising for OpenAI.
Product launches:
May 2024: Released GPT-4o ("o" for "omni"), a natively multimodal model processing text, images, and audio [28]
July 2024: Launched GPT-4o mini, a smaller, more efficient model [29]
September 2024: Introduced the o1 reasoning models (o1-preview and o1-mini), OpenAI's first models trained to use chain-of-thought reasoning before responding [30]
December 2024: Released Sora, the company's text-to-video generation model, as a standalone product for ChatGPT Plus and Pro subscribers [31]. Also launched o1 and o1-pro reasoning models with improved capabilities [32]
Key departures:
May 2024: Co-founder and Chief Scientist Ilya Sutskever departed to found Safe Superintelligence Inc. (SSI), a startup focused solely on building safe superintelligent AI. SSI raised $1 billion at launch and reached a $32 billion valuation by April 2025 [66]. Jakub Pachocki was appointed as new Chief Scientist. Superalignment co-lead Jan Leike also resigned, publicly stating that "safety culture and processes have taken a backseat to shiny products" [33]
August 2024: Co-founder John Schulman left to join Anthropic, citing a desire to deepen his focus on AI alignment and return to hands-on technical work. Schulman had led the reinforcement training organization that created ChatGPT. He left Anthropic less than a year later in early 2025 [67]. President Greg Brockman also took an extended leave of absence, returning in November 2024 to focus on leading OpenAI's infrastructure build-out [68]
September 2024: CTO Mira Murati announced her departure after 6.5 years at the company
October 2024: Miles Brundage, senior adviser for AGI readiness, resigned, writing that "neither OpenAI nor any other frontier lab is ready for AGI" [69]
Fundraising milestones:
October 2024: OpenAI raised $6.6 billion in its Series E round at a $157 billion valuation, the largest venture capital round in history at the time. The round was led by Thrive Capital (approximately $1.3 billion), with participation from Microsoft, Nvidia ($100 million), SoftBank ($500 million), Khosla Ventures, Altimeter Capital, and Fidelity. OpenAI also secured a $4 billion credit line from major banks including JPMorgan Chase, Citigroup, and Goldman Sachs [34]. Notably, the funding came with a condition: investors could claw back their money if OpenAI did not complete its conversion from non-profit to for-profit within two years [35]
Apple integration: In June 2024, Apple announced ChatGPT integration in iOS 18, macOS 15, and iPadOS 18
In January 2025, OpenAI announced the Stargate Project, a $500 billion AI infrastructure initiative with SoftBank, Oracle, and MGX, the largest private infrastructure investment in history [36]. The initial deployment committed $100 billion immediately, with a flagship 875-acre campus in Abilene, Texas.
On January 23, 2025, OpenAI launched Operator, a general-purpose AI agent that can autonomously control a web browser to perform tasks such as booking travel, making reservations, and shopping online [70]. Operator was powered by a new Computer-Using Agent (CUA) model that combined GPT-4o's vision capabilities with advanced reinforcement learning to interact with graphical user interfaces. Initially available only to ChatGPT Pro subscribers at operator.chatgpt.com, Operator was later integrated directly into ChatGPT as "ChatGPT agent" in July 2025 [71].
On July 17, 2025, OpenAI merged Operator and Deep Research into a single "ChatGPT agent" mode within ChatGPT [101]. Previously, Operator could scroll, click, and type on the web, while Deep Research excelled at analyzing and summarizing information, but they worked best in different situations. By integrating these complementary capabilities along with additional tools, OpenAI unlocked new functionality within one model. Users can select "agent mode" from the dropdown in the composer to access the combined capabilities. The standalone Operator site (operator.chatgpt.com) was subsequently sunset. ChatGPT agent became available starting July 18, 2025, for Pro, Plus, and Team subscribers, with Enterprise and Education access following in subsequent weeks [101].
In February 2025, OpenAI introduced Deep Research, an agentic capability within ChatGPT that conducts multi-step online research and synthesizes findings into comprehensive reports at the level of a research analyst [72]. Deep Research can browse and analyze hundreds of online sources over 5 to 30 minutes. Initially available only to Pro subscribers with a limit of 100 queries per month, access was later expanded in April 2025 to Plus, Team, Enterprise, and Edu users (25 queries/month) and free users (5 queries/month) [73]. The legacy Deep Research mode is scheduled for removal on March 26, 2026, as its functionality has been absorbed into the ChatGPT agent mode [116].
On October 21, 2025, OpenAI launched ChatGPT Atlas, an AI-native web browser built with ChatGPT at its core [102]. Atlas released initially for macOS, with versions for Windows, iOS, and Android announced for later availability. The browser is built on the OWL (OpenAI's Web Layer) architecture, which decouples Chromium's browser engine from the main app process using Mojo IPC, resulting in near-instant startup, improved stability, and a clean SwiftUI/AppKit codebase [103].
Key features of Atlas include:
In March 2026, OpenAI announced plans to merge ChatGPT, Codex, and the Atlas browser into a unified desktop "superapp," as reported by The Wall Street Journal based on an internal memo from Fidji Simo, OpenAI's CEO of Applications. The company plans first to equip Codex with additional agent-based features before integrating ChatGPT and Atlas into the consolidated application. The mobile ChatGPT app will remain unchanged [117].
In January 2025, Chinese AI startup DeepSeek released its R1 reasoning model, an open-source model that rivaled OpenAI's o1 at a fraction of the cost. DeepSeek reported training R1 in approximately two months for roughly $6 million, using only 2,788 Nvidia H800 GPUs, compared to the tens of thousands of GPUs and hundreds of millions of dollars OpenAI had spent on comparable models [74]. The release roiled financial markets, with Nvidia's stock losing 17% in a single day (the largest single-day wipeout in U.S. market history at the time) as investors questioned whether massive compute investments were truly necessary [75]. DeepSeek's app briefly topped the Apple App Store ahead of ChatGPT. The shock prompted OpenAI to accelerate its open-source strategy, contributing to its later decision to release the gpt-oss family of open-weight models [76].
March 2025: Raised $40 billion in a Series F round at a $300 billion valuation, led by SoftBank with participation from Microsoft and Coatue [37]
October 2025: Valuation reached $500 billion following a secondary share sale to SoftBank, Dragoneer, and Thrive Capital [7]
February 2026: Closed $110 billion funding round at a $730 billion pre-money valuation ($840 billion post-money), the largest private funding round in history. Amazon invested $50 billion (starting with a $15 billion initial commitment, followed by $35 billion when certain conditions are met), SoftBank pledged $30 billion in three equal tranches ($10 billion each on April 1, July 1, and October 1, 2026), and Nvidia committed $30 billion largely in the form of dedicated GPU compute capacity. Microsoft did not participate in this round but issued a joint statement confirming the existing partnership remains unchanged. As part of the deal, OpenAI committed to spending an additional $100 billion on Amazon Web Services over the next eight years and expanding collaboration with Nvidia, including 3 GW of dedicated inference capacity and 2 GW of training on Vera Rubin systems [100][104].
OpenAI's conversion from its original non-profit structure became one of the most closely watched corporate governance stories in the technology industry. In December 2024, the company announced plans to restructure so that the non-profit arm would no longer have full control over the for-profit entity. After significant public backlash, OpenAI modified its approach in May 2025, announcing that the for-profit entity would become a Public Benefit Corporation (PBC) while remaining under the non-profit's control [38].
In October 2025, OpenAI reached an agreement with the attorneys general of California and Delaware to finalize the restructuring. Under the new arrangement, the company was split into two entities: the OpenAI Foundation (a non-profit that holds approximately 26% of the stock in the for-profit, valued at roughly $130 billion, and appoints its board) and the OpenAI Group (a for-profit PBC free to raise funding and acquire companies) [2]. Microsoft holds a roughly 27% stake valued at about $135 billion. The restructuring drew criticism from those who argued it allowed the company to benefit unfairly from its earlier tax-exempt status and betrayed the expectations of original donors [39]. Following the February 2026 funding round, the OpenAI Foundation's stake increased in value to over $180 billion [100].
Notably, CEO Sam Altman did not receive any equity stake in the newly restructured company, despite earlier reports suggesting he could receive up to 7% (valued at roughly $10.5 billion). Altman's compensation remained a $76,001 annual salary with no ownership in OpenAI. His personal wealth (estimated at $3.1 billion by Forbes) derives from investments in other companies including Stripe, Reddit, and the nuclear fusion firm Helion Energy [77][78].
January 2025: Released o3-mini, a fast and efficient reasoning model, to all ChatGPT users including the free tier [40]
March 2025: Launched the Responses API and open-source Agents SDK, providing developers with new primitives for building agentic applications. The Responses API unified the best of the Chat Completions and Assistants APIs with built-in tools for web search, file search, computer use, and code interpretation [79]
April 2025: Released o3 and o4-mini reasoning models. These were the first OpenAI reasoning models capable of agentically combining every tool within ChatGPT (web search, file analysis, Python execution, image generation) and the first to "think with images," integrating visual information directly into the reasoning chain [41]
April 2025: Released GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano via the API. GPT-4.1 supported a 1 million-token context window and achieved a 21.4% improvement over GPT-4o on SWE-bench Verified. GPT-4.1 mini and nano offered cheaper, lighter options for high-volume workloads. GPT-4.1 later became available in ChatGPT in May 2025 [80]
May 2025: Launched Codex, a cloud-based software engineering agent powered by a specialized o3 variant (codex-1) [42]
June 2025: Released o3-pro for ChatGPT Pro and Team users, designed for high-reliability responses on challenging problems [43]
August 2025: Released GPT-5, a major generational leap featuring a unified architecture with a smart base model and a deeper reasoning model (GPT-5 thinking), connected by a real-time router. GPT-5 achieved 94.6% on AIME 2025 (without tools), 74.9% on SWE-bench Verified, 88% on Aider Polyglot, and 84.2% on MMMU. It also produced over five times fewer factual errors than o3. Priced at $1.25 per million input tokens and $10 per million output tokens, roughly half the input cost of GPT-4o. The API version launched with a 400,000-token context window [44]
August 2025: Released gpt-oss-120b and gpt-oss-20b, OpenAI's first open-weight language models since GPT-2 in 2019. Available under the Apache 2.0 license, gpt-oss-120b matched or exceeded o4-mini on core reasoning benchmarks while running on a single 80 GB GPU. gpt-oss-20b could run on edge devices with just 16 GB of memory. The release was partly a response to competitive pressure from DeepSeek and Meta's Llama models [81]
September 2025: Released Sora 2 with synchronized audio generation, identity-verified "cameos," and a TikTok-style social app on iOS and Android [45]. Five new Stargate datacenter sites were also announced, along with a strategic partnership with Nvidia for 10 gigawatts of GPU systems
December 2025: Released GPT-5.2 (December 11), the most capable model for professional knowledge work at that time, in three variants: Instant (optimized for speed and routine queries), Thinking (for complex structured work including coding, long documents, math, and planning), and Pro (maximum accuracy for difficult problems). GPT-5.2 Thinking produced outputs for knowledge work tasks at over 11x the speed and less than 1% the cost of expert professionals, and beat or tied top professionals on 70.9% of GDPval comparisons. GPT-5.2 Pro achieved 93.2% on GPQA Diamond, while Thinking set a new state of the art at 40.3% on FrontierMath [46]. One week later (December 18), released GPT-5.2-Codex, an agentic coding variant that scored 56.4% on SWE-Bench Pro and 64% on Terminal-Bench 2.0, with significantly stronger cybersecurity capabilities [47]
February 2026: On February 5, released GPT-5.3-Codex, the most capable agentic coding model to date, combining the frontier coding performance of GPT-5.2-Codex with the general reasoning capabilities of GPT-5.2. GPT-5.3-Codex scored 56.8% on SWE-Bench Pro, 77.3% on Terminal-Bench 2.0, and 64.7% on OSWorld-Verified, while running 25% faster than GPT-5.2-Codex. It was the first model that was instrumental in creating itself, as the Codex team used early versions to debug its own training and manage its own deployment. GPT-5.3-Codex supports interactive collaboration, allowing users to steer and interact with the model while it works without losing context [105][106]
March 2026: On March 3, released GPT-5.3 Instant, tuned for natural conversation with hallucination rates reduced by 26.8% when using the web and 19.7% when relying on internal knowledge. The model also significantly reduces unnecessary refusals and overly defensive or moralizing preambles [49][107]. On March 5, released GPT-5.4 (the first OpenAI model with native computer-use capabilities and a 1.05 million-token context window) and GPT-5.4 Pro. GPT-5.4 enables agents to execute UI-level commands such as file navigation, form completion, and spreadsheet editing across Windows and macOS. It scored 75.0% on OSWorld (exceeding the human baseline of 72.4%), 83% on GDPval (matching or exceeding professionals in 83% of comparisons), and 91% on BigLaw Bench. GPT-5.4 also features mid-response course correction, allowing users to adjust the model's plan while it works, and a new "tool search" system that lets models look up tool definitions as needed. GPT-5.4 Pro scored 89.3% on BrowseComp, 83.3% on ARC-AGI-2, and 38.0% on FrontierMath Tier 4 [48][108]. On March 17, released GPT-5.4 mini and GPT-5.4 nano, bringing many of GPT-5.4's strengths to faster, more efficient models; GPT-5.4 mini runs more than 2x faster than its predecessor and approaches GPT-5.4 performance on several evaluations [109]. As of March 11, GPT-5.1 models (Instant, Thinking, and Pro) were retired from ChatGPT [116].
On February 5, 2026, OpenAI launched Frontier, an enterprise platform for building, deploying, and managing AI agents across organizations [110]. Frontier acts as an intelligence layer that connects siloed data warehouses, CRM systems, ticketing tools, and internal applications, giving AI agents shared business context. The platform was designed around the metaphor of managing human employees: it offers agent onboarding, a feedback loop for continuous improvement, identity and access management (IAM) scoped to specific tasks, and enterprise-grade governance controls.
Key features of Frontier include:
| Feature | Description |
|---|---|
| Shared Business Context | Connects enterprise systems to create a semantic layer that all AI agents can reference |
| Agent Identity & IAM | Enterprise-grade identity management scoping access to exactly what each agent task requires |
| Agent Onboarding | Structured onboarding and review process for agents, similar to employee onboarding |
| Multi-Vendor Support | Compatible with OpenAI agents, custom-built agents, and third-party agents from Google, Microsoft, and Anthropic |
| Security & Compliance | SOC 2 Type II, ISO/IEC 27001, 27017, 27018, 27701, and CSA STAR certified |
Frontier initially launched to a limited set of customers including Uber, State Farm, Intuit, and Thermo Fisher Scientific, with broader availability planned over subsequent months [110][111].
On February 23, 2026, OpenAI announced multi-year partnerships with four of the world's largest consulting firms, forming the "Frontier Alliances" to help enterprises deploy AI agents at scale [112]. The partnerships are divided into two categories:
| Role | Partners | Focus |
|---|---|---|
| Strategy & Operating Model Partners | Boston Consulting Group (BCG), McKinsey & Company | Help leadership teams define where and how to deploy AI agents at scale |
| Systems Integration Partners | Accenture, Capgemini | End-to-end systems integration: data architecture, cloud infrastructure, and connecting Frontier to existing enterprise systems |
Each partner is investing in dedicated practice groups and building teams certified on OpenAI technology. The alliance partners work alongside OpenAI's Forward Deployed Engineering (FDE) team, combining OpenAI's research and product expertise with the consulting firms' deep transformation experience and global delivery capacity [112][113].
OpenAI significantly expanded its government footprint in early 2026. On February 27, 2026, hours after the Trump administration blacklisted Anthropic as a supply-chain risk to national security for refusing to allow its models to be used for mass surveillance or fully autonomous weapons, OpenAI announced a deal with the U.S. Department of Defense (renamed "Department of War" by the administration) to deploy its AI models in classified military environments [114]. The contract, with a $200 million ceiling, covers prototyping how frontier AI can improve administrative operations including healthcare for service members, streamlining acquisition data, and supporting cyber defense.
The deal drew immediate backlash, and CEO Sam Altman admitted the rollout appeared "opportunistic and sloppy." On March 2, 2026, OpenAI amended the contract, explicitly prohibiting use of its technology for domestic surveillance of U.S. persons and nationals and barring use by Department of War intelligence agencies such as the NSA. The amendment also expanded protections to cover "commercially acquired" public data, closing a loophole that would have left geolocation, web browsing data, and personal financial information purchased from data brokers unprotected [114][115].
On March 17, 2026, OpenAI expanded its government reach further by partnering with Amazon Web Services (AWS) to distribute its AI products across U.S. government agencies for both classified and unclassified work through AWS GovCloud and AWS Classified Regions. OpenAI retains control over which models are made available, and AWS must provide notice before enabling access for especially sensitive agencies, including intelligence customers [118].
OpenAI announced plans to launch an AI-powered jobs platform by mid-2026, directly competing with LinkedIn and Indeed in the professional networking and recruitment space [119]. The platform will use AI algorithms to match candidates with employers based on demonstrated AI competencies rather than traditional resume keywords. OpenAI is also developing AI Certifications through OpenAI Academy in collaboration with major employers including Walmart and John Deere, with credentials ranging from basic workplace AI literacy to advanced skills such as prompt engineering. OpenAI aims to certify at least 10 million Americans in workplace AI skills by 2030 [119].
May 2025: Acquired Jony Ive's hardware startup "io" for $6.5 billion, OpenAI's largest acquisition. Ive became Chief Design Officer, leading OpenAI's consumer hardware product development. The deal closed in July 2025 [82]
May 2025: Announced a $3 billion acquisition of Windsurf (formerly Codeium), an AI coding tool. However, the deal collapsed in July 2025 due to a disagreement with Microsoft over intellectual property rights. Microsoft expected its broad IP rights over OpenAI technology to extend to Windsurf's assets, but OpenAI refused, viewing Windsurf as a competitive advantage against Microsoft's own GitHub Copilot. Google DeepMind subsequently hired Windsurf's CEO and key researchers in a $2.4 billion licensing and talent deal [83]
May 2025: Acquired product development platform Statsig for $1.1 billion
By mid-2025, 92% of Fortune 500 companies were using ChatGPT, with over 1 million organizations using OpenAI's technology by December 2025 [50]. The business user base grew to over 9 million paying business users by February 2026 [51]
OpenAI's relationship with AI safety continued to generate controversy through 2025 and into 2026. In September 2024, the company created a new Mission Alignment team led by Joshua Achiam to promote OpenAI's stated mission. However, in February 2026, OpenAI disbanded the Mission Alignment team after just 16 months, transferring its seven employees to other teams. Achiam took a new title as "chief futurist" [84]. The disbanding followed the earlier dissolution of the Superalignment team in May 2024 and drew criticism as part of a broader pattern. Observers noted that by early 2026, essentially none of the people most associated with AI safety at OpenAI (Sutskever, Leike, Murati, Schulman, Brundage, and others) remained in positions of influence [85].
Separately, OpenAI removed the word "safely" from its corporate mission statement during this period, prompting further questions about whether the company's priorities had shifted away from cautious AI development [60].
OpenAI's revenue growth accelerated sharply. The company generated $3.7 billion in full-year 2024 revenue, then exceeded that figure in just the first half of 2025 ($4.3 billion in H1 2025). Full-year 2025 revenue came in at $13.1 billion [52][86]. By late 2025, the annualized revenue run rate surpassed $20 billion, according to CFO Sarah Friar [53]. By the end of February 2026, the annualized revenue run rate topped $25 billion, driven by surging enterprise adoption and growing consumer subscriptions [120]. However, the company's burn rate remains substantial (approximately $700 million per month in mid-2025), and profitability is not expected before 2029 or 2030 [54].
In January 2025, Sam Altman published a blog post stating that "we are now confident we know how to build AGI as we have traditionally understood it" and predicted that "in 2025, we may see the first AI agents join the workforce and materially change the output of companies" [87]. He also signaled that OpenAI's ambitions extended beyond AGI to artificial superintelligence (ASI), writing: "We are beginning to turn our aim beyond that to superintelligence in the true sense of the word" [88]. However, Altman also acknowledged that "AGI" had become a "sloppy term," and after speculation that OpenAI's o3-mini model had achieved AGI, he pushed back, posting: "We are not gonna deploy AGI next month, nor have we built it" [89].
Following the October 2025 restructuring, OpenAI operates under a new corporate structure [2][38]:
| Entity | Type | Role | Established |
|---|---|---|---|
| OpenAI Foundation | Non-profit | Holds | 2025 (successor to OpenAI Inc.) |
| OpenAI Group | Public Benefit Corporation | Commercial operations; product development; employee equity; free to raise funding and acquire companies | 2025 (successor to OpenAI Global LLC) |
The earlier structure consisted of OpenAI Inc. (a 501(c)(3) non-profit), OpenAI Global LLC (a capped-profit subsidiary created in 2019), OpenAI Holdings LLC (an investment vehicle), and OpenAI GP LLC (a general partner entity). Under the original capped-profit model, investor returns were limited to 100x their investment, with excess value flowing to the non-profit [14].
Under the final restructuring, Microsoft holds approximately 27% of the for-profit entity (valued at ~$135 billion), the OpenAI Foundation holds approximately 26%, and the remaining ~47% is held by other investors and employees [77].
| Name | Role | Background | Joined |
|---|---|---|---|
| Bret Taylor | Chairman | Former co-CEO of Salesforce, former CTO of Facebook | November 2023 |
| Sam Altman | CEO, Director | Co-founder, former president of Y Combinator | March 2024 (rejoined board) |
| Adam D'Angelo | Director | CEO of Quora | October 2023 |
| Larry Summers | Director | Former U.S. Treasury Secretary, former Harvard president | November 2023 |
| Dr. Sue Desmond-Hellmann | Director | Former CEO of Bill & Melinda Gates Foundation | March 2024 |
| Nicole Seligman | Director | Former EVP and General Counsel of Sony Corporation | March 2024 |
| Fidji Simo | Director | CEO of Instacart | March 2024 |
| Paul Nakasone | Director | Former NSA Director, retired U.S. Army General | June 2024 |
| Zico Kolter | Director | Carnegie Mellon professor, AI safety researcher | November 2024 |
| Microsoft | Observer (non-voting) | Strategic partner and major investor | 2023 |
| Position | Name | Tenure | Previous Role |
|---|---|---|---|
| Chief Executive Officer | Sam Altman | 2019-2023, 2023-present | President of Y Combinator |
| President | Greg Brockman | 2015-present | Co-founder, former CTO of Stripe |
| Chief Financial Officer | Sarah Friar | 2024-present | CEO of Nextdoor |
| Chief Operating Officer | Brad Lightcap | 2018-present | VP at Y Combinator |
| Chief Scientist | Jakub Pachocki | 2024-present | Research lead at OpenAI |
| Chief Product Officer | Kevin Weil | 2024-present | Former VP at Instagram |
| Chief Design Officer | Jony Ive | 2025-present | Former Chief Design Officer at Apple; co-founded io |
OpenAI has experienced exponential revenue growth [52]:
| Year | Annual Revenue | Growth Rate | Key Drivers |
|---|---|---|---|
| 2020 | $3.5 million | - | Early API access |
| 2021 | $28 million | 700% | GPT-3 API expansion |
| 2022 | $300 million | 971% | Increased enterprise adoption |
| 2023 | $1.6 billion | 433% | ChatGPT launch, ChatGPT Plus |
| 2024 | $3.7 billion | 131% | Enterprise subscriptions, API growth |
| 2025 | $13.1 billion | 254% | New products, enterprise expansion, API growth |
| 2026 (projected) | $29-30 billion | ~125% | Frontier platform, enterprise agents, hardware products |
| 2029 (projected) | $100 billion | - | AGI commercialization target |
Individual ChatGPT Plus subscribers represented the largest share of revenue as of mid-2024, far outpacing both enterprise products and API revenue [55]. By 2025, enterprise and API revenue had grown substantially as a proportion of the total.
OpenAI's average stock-based compensation reached $1.5 million per employee in 2025, the highest of any major tech startup. Roughly 46% of the company's annual revenue goes toward stock-based compensation, reflecting the intense competition for AI talent [90].
OpenAI has raised approximately $168 billion across multiple funding rounds [52]:
| Date | Round | Amount | Valuation | Lead Investors | Key Terms |
|---|---|---|---|---|---|
| December 2015 | Seed | $1 billion (pledged) | - | Founders, Elon Musk, Peter Thiel, Reid Hoffman | Non-profit donations |
| July 2019 | Series A | $1 billion | $1 billion | Microsoft | Capped-profit structure introduced |
| January 2023 | Extended Series A | $10 billion | $29 billion | Microsoft | Revenue sharing agreement |
| April 2024 | Series D | $300 million | $80 billion | Tiger Global, Sequoia | Employee tender offer |
| October 2024 | Series E | $6.6 billion | $157 billion | Thrive Capital, Microsoft, Nvidia, SoftBank | Largest VC round at the time; $4B credit line; for-profit conversion clause [34] |
| March 2025 | Series F | $40 billion | $300 billion | SoftBank, Microsoft, Coatue | Debt and equity mix |
| October 2025 | Secondary Sale | $6.6 billion | $500 billion | SoftBank, Dragoneer, Thrive | Employee liquidity event |
| February 2026 | Series G | $110 billion | $730B pre / $840B post | Amazon ($50B), SoftBank ($30B), Nvidia ($30B) | Largest private round ever; Amazon cash + Nvidia compute; AWS expansion [100] |
Notable investors include:
Microsoft: $14+ billion total investment
Amazon: $50 billion (February 2026)
SoftBank: $60+ billion (including Stargate and Series G)
Nvidia: $30+ billion (largely GPU compute)
Sequoia Capital, Andreessen Horowitz, Khosla Ventures
| Metric | Value | Date |
|---|---|---|
| Annualized Revenue Run Rate | $25+ billion | February 2026 |
| Monthly Burn Rate | $700 million | Mid-2025 |
| R&D Expenses (H1 2025) | $6.7 billion | June 2025 |
| Total Employees | ~4,000+ | Early 2026 |
| Compute Infrastructure Costs | ~50% of expenses | 2025 |
| Projected Break-even | 2029-2030 | Per investor documents |
| Total Cash Burn (projected through 2029) | $115 billion | Per financial models |
| Projected Compute Spending (through 2030) | ~$600 billion | Per February 2026 investor update |
| Weekly Codex Users | 1.6 million | February 2026 |
The Generative Pre-trained Transformer (GPT) series represents OpenAI's flagship language models:
| Model | Release Date | Parameters | Context Length | Key Features | Training Data |
|---|---|---|---|---|---|
| GPT-1 | June 2018 | 117 million | 512 tokens | First transformer-based LM, unsupervised pre-training | BookCorpus (7,000 books) |
| GPT-2 | February 2019 | 1.5 billion | 1,024 tokens | Zero-shot task transfer, initially withheld | 40GB WebText |
| GPT-3 | June 2020 | 175 billion | 2,048 tokens | Few-shot learning, API commercialization | 570GB diverse text |
| GPT-3.5 | November 2022 | 175 billion | 4,096 tokens | RLHF fine-tuning, ChatGPT base | Updated dataset + human feedback |
| GPT-4 | March 2023 | ~1.76 trillion (est.) | 8,192/32,768 tokens | Multimodal (text+vision), improved reasoning | Undisclosed |
| GPT-4 Turbo | November 2023 | Undisclosed | 128,000 tokens | Cost reduction, JSON mode, improved instruction following | Training cutoff: April 2023 |
| GPT-4o | May 2024 | Undisclosed | 128,000 tokens | Native multimodal (text+vision+audio), 2x speed, 50% cost reduction | Training cutoff: October 2023 |
| GPT-4o mini | July 2024 | Undisclosed | 128,000 tokens | 60% cheaper than GPT-3.5, optimized for high-volume tasks | Training cutoff: October 2023 |
| GPT-4.1 | April 14, 2025 | Undisclosed | 1,000,000 tokens | 21.4% improvement over GPT-4o on SWE-bench; strong instruction following | Training cutoff: June 2024 |
| GPT-4.1 mini | April 14, 2025 | Undisclosed | 1,000,000 tokens | Cost-efficient replacement for GPT-4o mini | Training cutoff: June 2024 |
| GPT-4.1 nano | April 14, 2025 | Undisclosed | 1,000,000 tokens | Lightest model for edge and high-volume use | Training cutoff: June 2024 |
| GPT-5 | August 7, 2025 | Undisclosed | 400,000 tokens (API) | Unified routing system with base and thinking modes; 94.6% AIME 2025; 74.9% SWE-bench; 5x fewer hallucinations than o3 | Current web data |
| GPT-5.2 | December 11, 2025 | Undisclosed | 400,000 tokens | Three variants (Instant, Thinking, Pro); 93.2% GPQA Diamond (Pro); 40.3% FrontierMath (Thinking); 70.9% professional parity | Training cutoff: Aug 31, 2025 |
| GPT-5.2-Codex | December 18, 2025 | Undisclosed | 400,000 tokens | Agentic coding; 56.4% SWE-Bench Pro; 64% Terminal-Bench 2.0; enhanced cybersecurity | Optimized for code |
| GPT-5.3-Codex | February 5, 2026 | Undisclosed | 400,000 tokens | Agentic coding + general reasoning; 56.8% SWE-Bench Pro; 77.3% Terminal-Bench 2.0; 64.7% OSWorld-Verified; 25% faster; interactive collaboration | Optimized for code + professional work |
| GPT-5.3 Instant | March 3, 2026 | Undisclosed | 400,000 tokens | Natural conversation; hallucination reduced 26.8% (web), 19.7% (internal); reduced unnecessary refusals | Updated dataset |
| GPT-5.4 | March 5, 2026 | Undisclosed | 1,050,000 tokens | Native computer-use; tool search; 75.0% OSWorld; 83% GDPval; 91% BigLaw Bench; 33% fewer factual errors; mid-response course correction | Current web data |
| GPT-5.4 Pro | March 5, 2026 | Undisclosed | 1,050,000 tokens | Highest-capability variant; 89.3% BrowseComp; 83.3% ARC-AGI-2; 38.0% FrontierMath Tier 4 | Current web data |
| GPT-5.4 mini | March 17, 2026 | Undisclosed | Undisclosed | 2x faster than predecessor; approaches GPT-5.4 on SWE-Bench Pro and OSWorld-Verified | Current web data |
| GPT-5.4 nano | March 17, 2026 | Undisclosed | Undisclosed | Lightest GPT-5.4 variant for edge and high-volume use | Current web data |
In August 2025, OpenAI released its first open-weight models since GPT-2, marking a significant shift in strategy partly driven by competitive pressure from DeepSeek and Meta's Llama models [81]:
| Model | Parameters | License | Key Performance | Hardware Requirements |
|---|---|---|---|---|
| gpt-oss-120b | 120 billion | Apache 2.0 | Matches or exceeds o4-mini on reasoning benchmarks; outperforms o4-mini on health and competition math | Single 80 GB GPU |
| gpt-oss-20b | 20 billion | Apache 2.0 | Similar to o3-mini on common benchmarks | Edge devices with 16 GB memory |
Both models are available for download on Hugging Face and GitHub, and can be run locally through LM Studio and Ollama.
The o-series represents OpenAI's approach to chain-of-thought reasoning, using "reasoning tokens" to think through problems before responding. These models are trained with reinforcement learning to refine their internal chains of thought, recognizing and correcting mistakes and trying different approaches when needed [30]:
| Model | Release | Specialization | Performance | Availability |
|---|---|---|---|---|
| o1-preview | September 12, 2024 | General reasoning | 83rd percentile on Codeforces | ChatGPT Plus/Pro |
| o1-mini | September 12, 2024 | STEM reasoning | 70th percentile on AIME | ChatGPT Plus/Pro, API |
| o1 | December 2024 | Advanced reasoning | 89th percentile on Codeforces; PhD-level on GPQA Diamond | ChatGPT Pro, API |
| o1-pro | December 2024 | Maximum reliability | Extended thinking for hardest problems | ChatGPT Pro only |
| o3-mini | January 31, 2025 | Fast, efficient reasoning | Strong STEM, math, and coding performance | All ChatGPT users, API [40] |
| o3 | April 16, 2025 | Next-gen reasoning | 88.9% AIME 2025; 20% fewer major errors than o1 on real-world tasks | ChatGPT Plus/Pro, API [41] |
| o4-mini | April 16, 2025 | Cost-efficient reasoning | 92.7% AIME 2025; first model to "think with images" | All ChatGPT users, API [41] |
| o3-pro | June 10, 2025 | High-reliability reasoning | Preferred over o3 in every tested category by expert reviewers | ChatGPT Pro/Team, API [43] |
ChatGPT is OpenAI's conversational AI interface launched November 30, 2022 [19]:
| Tier | Monthly Cost | Features | Model Access | Usage Limits |
|---|---|---|---|---|
| Free | $0 | Basic access | GPT-4o (limited), GPT-3.5 | Limited daily messages |
| Go | $8 | Expanded GPT-5.2 Instant access, file uploads, image creation | GPT-5.2 Instant | 10x more messages than Free |
| Plus | $20 | Priority access, faster response, Deep Research (25/month) | GPT-5, o-series, DALL-E 3, Codex | Higher limits on advanced models |
| Pro | $200 | Maximum access, Deep Research (250/month) | All models including o3-pro | Unlimited on most models |
| Business | $25/user (annual) | Collaborative workspace, admin tools, credits system | All Plus models | Higher limits, shared workspace |
| Enterprise | Custom | SSO, data privacy, admin controls, unlimited GPT-5 | All models, custom fine-tuning | Unlimited, SLA guarantees |
Note: The "Team" plan was renamed to "Business" in August 2025. The "Go" plan launched globally in January 2026 after a pilot in India (August 2025) [91][92].
Usage Statistics:
| Milestone | Weekly Active Users | Date |
|---|---|---|
| ChatGPT launch | N/A (100M total users in 2 months) | November 2022 |
| Early growth | 300 million | December 2024 |
| Rapid acceleration | 400 million | February 2025 |
| Continued growth | 500 million | March 2025 |
| Pre-GPT-5 surge | 700 million | August 2025 |
| Post-GPT-5 | 800 million | October 2025 |
| Current | 900 million | February 2026 |
As of early 2026, ChatGPT is available in 185 countries and 50+ languages, with 92% of Fortune 500 companies using the product, over 9 million paying business users, and more than 500 million mobile app downloads across iOS and Android [5][50][51].
Key Features:
Custom GPTs: User-created specialized chatbots
Code Interpreter: Python execution environment
Web browsing: Real-time internet access
Voice conversations: Natural speech interaction
Vision: Image understanding and analysis
Memory: Contextual information retention across chats
Canvas: Collaborative editing interface
Deep Research: Multi-step agentic research reports
Operator / ChatGPT Agent: Autonomous web browser control
Atlas Browser: AI-native web browsing with contextual assistance
DALL-E is OpenAI's text-to-image generation system:
| Version | Release | Resolution | Key Features | Availability |
|---|---|---|---|---|
| DALL-E | Jan 2021 | 256x256 | 12B parameters, first text-to-image transformer | Research preview |
| DALL-E 2 | Apr 2022 | 1024x1024 | 3.5B parameters, inpainting, variations | API, Labs interface |
| DALL-E 3 | Sep 2023 | 1024x1024, 1792x1024, 1024x1792 | ChatGPT integration, improved text rendering | ChatGPT, API |
| GPT-4o Image | Mar 2025 | Up to 4K | Replaced DALL-E 3, faster generation, style consistency | ChatGPT exclusive |
Advanced Features:
Inpainting: Edit specific regions of existing images
Outpainting: Extend images beyond original boundaries
Variations: Generate similar images from a source
Style transfer: Apply artistic styles to images
C2PA metadata: Content authentication standard
Sora is OpenAI's text-to-video generation technology. The model was first announced on February 15, 2024, with sample videos that generated significant public interest, though it was not made available to the public at that time [56]. The consumer launch occurred on December 9, 2024, when OpenAI released Sora Turbo for ChatGPT Plus and Pro subscribers in the US and initially limited regions [31].
| Version | Release | Capabilities | Max Duration | Resolution |
|---|---|---|---|---|
| Sora (preview) | February 2024 (announced) | Text-to-video research preview | 60 seconds (demo) | Up to 1080p |
| Sora Turbo | December 9, 2024 | Text-to-video, image animation, remixing | 20 seconds | Up to 1080p |
| Sora 2 | September 2025 | Video+audio, identity-verified "cameos", social features | 60 seconds | Up to 4K |
Sora 2 Features:
Synchronized dialogue and sound effects
iOS/Android apps with TikTok-style feed
"Cameos": Identity-verified appearances in generated videos
Visible and invisible watermarks for authenticity
Multiple aspect ratios and styles
Character consistency across scenes
Limitations: At launch, OpenAI acknowledged that Sora still struggled with unrealistic physics and complex actions over long durations [31].
| Product | Launch | Description | Capabilities | Pricing |
|---|---|---|---|---|
| OpenAI Codex | Aug 2021 | Code-trained GPT-3 variant | Code completion, translation between languages | Deprecated, succeeded by GPT-4 |
| GitHub Copilot | Jun 2021 | AI pair programmer | Real-time code suggestions, whole function generation | $10/month individual, $19/month business |
| Codex (2025) | May 2025 | Cloud software engineering agent | Autonomous coding, parallel task execution, full project development | ChatGPT Pro, API access |
| GPT-5.2-Codex | December 18, 2025 | Agentic coding model | Large refactors, migrations, cybersecurity; 56.4% SWE-Bench Pro | ChatGPT paid tiers, API [47] |
| GPT-5.3-Codex | February 5, 2026 | Advanced agentic coding + general reasoning | 56.8% SWE-Bench Pro; 77.3% Terminal-Bench 2.0; 64.7% OSWorld; interactive collaboration; 25% faster | ChatGPT paid tiers, Codex app, CLI, IDE extension [105] |
Codex (2025) Features:
Cloud-based development environment
Integration with GitHub, VS Code, Cursor, Windsurf
Powered by codex-1 model (optimized o3 variant)
Parallel execution of multiple coding tasks
Automated testing and debugging
Full-stack application development
GPT-5.3-Codex Advances (February 2026):
Combines frontier coding performance with general reasoning and professional knowledge in one model
Supports the full software lifecycle: debugging, deploying, monitoring, writing PRDs, editing copy, user research, tests, and metrics
Interactive collaboration: users can steer and interact with the model while it works without losing context
First model that was instrumental in creating itself (used to debug its own training and manage its own deployment) [105]
Weekly Codex users tripled since the start of 2026 to 1.6 million [100]
OpenAI's API platform has grown into one of the most widely used AI developer ecosystems. Key milestones include:
| Date | Development | Description |
|---|---|---|
| June 2020 | GPT-3 API launch | First commercial LLM API |
| November 2023 | Assistants API | Stateful, multi-turn agent framework with threads and tool use |
| March 2025 | Responses API | Unified API primitive combining Chat Completions and Assistants APIs; includes built-in tools for web search, file search, computer use, and code interpreter [79] |
| March 2025 | Agents SDK | Open-source SDK (Python, TypeScript) for building multi-agent workflows with tool use, handoffs, guardrails, and tracing; provider-agnostic [79] |
| 2025 | Fine-tuning API | Custom model fine-tuning for GPT-4o and other models |
| 2025 | Batch API | 50% cost reduction for non-real-time workloads |
The Responses API is recommended for all new projects as of 2025, though Chat Completions remains supported. The API supports over 100 languages, structured JSON output, function calling, and streaming responses.
Whisper is an automatic speech recognition system [57]:
| Model Size | Parameters | Relative Speed | English WER | Multilingual Support |
|---|---|---|---|---|
| tiny | 39M | ~32x | 7.5% | Limited |
| base | 74M | ~16x | 5.0% | Good |
| small | 244M | ~6x | 3.5% | Good |
| medium | 769M | ~2x | 2.9% | Better |
| large-v3 | 1.55B | 1x | 2.5% | Best (50+ languages) |
Capabilities:
Trained on 680,000 hours of multilingual data
Automatic language detection
Translation to English
Timestamp generation
Speaker diarization (with additional processing)
Open-source release
| Platform | Status | Purpose | Key Features |
|---|---|---|---|
| OpenAI Gym | Deprecated (2022) | RL benchmarking | Standard environments, common API, leaderboards |
| Gymnasium | Active (community fork) | RL research | Maintained by Farama Foundation, backward compatible |
| OpenAI Universe | Deprecated (2017) | General AI training | Browser/app control, pixel-based interaction |
CLIP (Contrastive Language-Image Pre-training) connects text and images [58]:
Trained on 400 million image-text pairs
Zero-shot image classification
Powers many multimodal applications
Open-source release (MIT license)
Research project for improved factual accuracy:
Web browsing during inference
Citation of sources
Fact-checking capabilities
Influenced ChatGPT's web browsing feature
Announced January 21, 2025, the Stargate Project represents the largest private infrastructure investment in history [36]:
| Aspect | Details |
|---|---|
| Total Investment | $500 billion over 4 years |
| Initial Deployment | $100 billion immediate (2025) |
| Equity Partners | SoftBank (lead), OpenAI, Oracle, MGX (UAE) |
| Technology Partners | Arm, Microsoft, Nvidia, Oracle |
| Leadership | Masayoshi Son (Chairman), Sam Altman (Operations) |
| Primary Locations | Abilene, Texas (operational); 5 additional sites (Sep 2025) |
| Secondary Locations | New Mexico, Ohio, undisclosed Midwest sites |
| Power Capacity Goal | 10 gigawatts by end of 2025 |
| Current Capacity | ~7 gigawatts across all sites |
| Job Creation Target | 100,000+ jobs |
| Computing Power | Equivalent to millions of GPUs |
Stargate Facilities:
Abilene, Texas: 875-acre flagship campus, 1M sq ft datacenter
Oracle Cloud Region: Dedicated OpenAI infrastructure
Geographic distribution: Designed for redundancy and latency optimization
Renewable energy: Commitment to sustainable power sources
Greg Brockman has taken a leading role in Stargate's execution, serving as the behind-the-scenes architect translating Altman's infrastructure vision into hardware, investment, and partnerships [68].
The Microsoft-OpenAI partnership has evolved significantly [59]:
| Period | Investment | Key Terms | Products |
|---|---|---|---|
| 2019-2022 | $1 billion | Azure exclusivity, joint development | Azure OpenAI Service |
| 2023 | $10 billion | Revenue sharing (25% to Microsoft), compute credits | Bing Chat, Microsoft 365 Copilot |
| 2024 | Continued | IP licensing for Microsoft products | Windows Copilot, GitHub Copilot |
| 2025 | $14+ billion total | Right of first refusal on capacity (not exclusive), bidirectional agreements; ~27% equity stake in restructured OpenAI Group | Expanded Azure AI offerings |
Current Partnership Terms (2025):
Microsoft retains IP rights to pre-AGI OpenAI technology
OpenAI API remains exclusive to Azure through 2030
Revenue sharing modified but continues
Microsoft has board observer status (non-voting)
Microsoft holds approximately 27% equity in the for-profit entity (~$135 billion)
Collaboration on AI safety and research continues
The partnership experienced some friction in 2025. The failed Windsurf acquisition highlighted tensions over IP rights: Microsoft expected its broad licensing agreement to extend to acquired companies' technology, while OpenAI resisted, viewing the Windsurf deal as a means to compete against Microsoft's own GitHub Copilot [83]. Notably, Microsoft did not participate in OpenAI's $110 billion February 2026 funding round, though both companies stated the existing partnership remains unchanged [100].
| Partner | Type | Announcement | Details |
|---|---|---|---|
| Apple | Integration | June 2024 | ChatGPT integration in iOS 18, macOS 15, iPadOS 18 |
| Nvidia | Infrastructure | Sep 2025 | 10 gigawatts of GPU systems for Stargate |
| Oracle | Infrastructure | Jan 2025 | Stargate co-investor, dedicated cloud region |
| SoftBank | Investment/Infrastructure | 2024-2026 | $60+ billion investment, Stargate leadership |
| Amazon | Cloud Services / Investment | Feb 2026 | $50 billion investment; $100 billion AWS commitment over 8 years; Bedrock integration [100] |
| SAP | Enterprise | Sep 2025 | "OpenAI for Germany" sovereign cloud project |
| AWS | Government | Mar 2026 | Distribution of OpenAI products to U.S. government agencies via GovCloud and Classified Regions [118] |
| Accenture | Frontier Alliance | Feb 2026 | Systems integration partner for Frontier platform deployment [112] |
| Capgemini | Frontier Alliance | Feb 2026 | Systems integration partner for Frontier platform deployment [112] |
| Boston Consulting Group | Frontier Alliance | Feb 2026 | Strategy and operating model partner for Frontier platform [112] |
| McKinsey & Company | Frontier Alliance | Feb 2026 | Strategy and operating model partner for Frontier platform [112] |
OpenAI operates according to its Charter principles, published in 2018 [3]:
Broadly Distributed Benefits: Commit to use any influence over AGI's deployment to ensure it benefits all of humanity, avoiding uses that harm humanity or concentrate power
Long-Term Safety: Conduct research to make AGI safe and encourage adoption of safety measures across the AI community; assist value-aligned, safety-conscious projects that come close to building AGI before OpenAI
Technical Leadership: Maintain position at the forefront of AI capabilities to effectively address AGI's impact on society, policy and safety advocacy alone insufficient without technical expertise
Cooperative Orientation: Actively cooperate with other research and policy institutions; create a global community working together to address AGI's global challenges
Notably, following the 2025 restructuring, OpenAI removed the word "safely" from its corporate mission statement, raising questions about whether the company's priorities had shifted [60].
OpenAI maintains multiple safety initiatives, though the organizational structure has changed significantly over 2024-2026:
| Initiative | Description | Status |
|---|---|---|
| Preparedness Framework | Systematic evaluation of model risks across domains | Active, quarterly updates |
| Red teaming | Adversarial testing with domain experts | Ongoing for all major releases |
| Constitutional AI | Principles for AI behavior and alignment | Integrated in training |
| Superalignment team | Long-term AGI alignment research | Dissolved May 2024, integrated into other teams |
| Mission Alignment team | Promoting OpenAI's mission to benefit humanity | Created September 2024, dissolved February 2026 [84] |
| Safety evaluations | Assessment across cybersecurity, CBRN, persuasion, autonomy | Required pre-deployment |
| Content filtering | Multi-layered harmful content prevention | Continuously updated |
OpenAI's Usage Policies (updated October 17, 2025) establish guidelines for responsible AI use [61]:
Key Policy Principles:
User Empowerment: Enable innovation while maintaining safety
Responsible Use: Users accountable for appropriate usage
Safety First: Prioritize safety in monitoring and enforcement
Evolving Rules: Adapt policies to emerging use cases
Major Prohibitions:
Threats to persons or property (harassment, violence, weapons)
Privacy violations (unauthorized data collection, surveillance)
Child safety risks (CSAM, grooming, inappropriate content)
Automated high-stakes decisions without human review
Misinformation campaigns or electoral manipulation
Academic dishonesty or professional malpractice
| Year | Innovation | Impact | Citation Count |
|---|---|---|---|
| 2017 | Proximal Policy Optimization (PPO) | Standard RL algorithm | 10,000+ |
| 2018 | GPT: Improving Language Understanding | Launched transformer era in NLP | 15,000+ |
| 2019 | Sparse Transformer | Efficient attention mechanisms | 3,000+ |
| 2020 | GPT-3: Language Models are Few-Shot Learners | Demonstrated emergent abilities | 20,000+ |
| 2021 | CLIP: Connecting Text and Images | Multimodal understanding | 8,000+ |
| 2022 | InstructGPT: Training with Human Feedback | RLHF methodology | 5,000+ |
| 2023 | GPT-4 Technical Report | Multimodal reasoning | 7,000+ |
| 2024 | Direct Preference Optimization | Simplified alignment training | 2,000+ |
Whisper: Complete ASR system (Apache 2.0 license)
CLIP: Vision-language model (MIT license)
Triton: GPU programming language (MIT license)
Gym: RL environments (MIT license, now community-maintained)
Baselines: RL algorithm implementations
GPT-2: Model weights and code (custom license)
gpt-oss-120b / gpt-oss-20b: Open-weight reasoning models (Apache 2.0 license, August 2025)
Agents SDK: Multi-agent framework (open-source, March 2025)
OpenAI faces multiple lawsuits regarding use of copyrighted material:
| Plaintiff | Filing Date | Claims | Status |
|---|---|---|---|
| Authors Guild + 17 authors | Sep 2023 | Copyright infringement in training | Ongoing |
| The New York Times | Dec 2023 | Unlicensed use of articles, unfair competition | Ongoing, seeking billions in damages |
| Sarah Silverman et al. | Jul 2023 | Book piracy for training data | Partially dismissed, ongoing |
| Getty Images | Feb 2023 | Image copyright violations | Ongoing |
| Various artists | 2023-2024 | DALL-E training on copyrighted art | Class action forming |
OpenAI's defense centers on fair use doctrine and transformative nature of AI systems [62].
Elon Musk has filed multiple lawsuits against OpenAI (2024-2026):
| Date | Claim | OpenAI Response | Status |
|---|---|---|---|
| Feb 2024 | Breach of founding agreement to remain non-profit | Musk knew of for-profit plans | Withdrawn |
| Aug 2024 | Antitrust violations | Baseless claims | Dismissed |
| Oct 2025 | xAI trade secret theft via employee poaching | "Harassment strategy," denied | Dismissed (Feb 2026) |
| 2025 | Fraud claims seeking $79-134 billion in damages | Unreliable damages methodology | Trial scheduled April-May 2026 |
The most significant remaining case alleges that OpenAI co-founders Sam Altman and Greg Brockman defrauded Musk when he donated $38 million in seed funding by misrepresenting OpenAI's commitment to remaining a non-profit. Musk's expert witness calculated potential damages of $79 billion to $134 billion in ill-gotten profits. A jury trial is scheduled to begin on April 27, 2026 and run through May [93][94]. A separate xAI trade secret lawsuit (alleging that former xAI engineer Xuechen Li downloaded xAI's "entire codebase" before joining OpenAI) was dismissed by a federal judge in February 2026 [95].
Musk has stated he would donate any lawsuit winnings to charity [96].
Key Departures and Criticisms:
May 2024: Ilya Sutskever (co-founder, Chief Scientist) departed to found Safe Superintelligence Inc. Jan Leike (Superalignment co-lead) resigned, publicly stating that "safety culture and processes have taken a backseat to shiny products" [33]. The Superalignment team was dissolved, with its work distributed across other teams.
August 2024: Co-founder John Schulman left for Anthropic, citing a desire for deeper focus on AI alignment [67]. He later left Anthropic as well in early 2025.
Sep 2024: Multiple safety researchers left for Anthropic and other safety-focused organizations
October 2024: Miles Brundage, senior adviser for AGI readiness, resigned, stating "neither OpenAI nor any other frontier lab is ready for AGI" [69]
February 2026: Mission Alignment team disbanded after 16 months, its seven members reassigned [84]
2025-2026: OpenAI removed the word "safely" from its mission statement [60]
By early 2026, essentially all of the researchers and executives most associated with AI safety at OpenAI had departed, raising concerns among AI safety researchers and policymakers about the company's long-term commitment to responsible development [85].
OpenAI's February 2026 deal with the U.S. Department of Defense drew significant criticism, particularly because it was announced hours after the Trump administration blacklisted Anthropic for refusing to allow its technology to be used for mass surveillance or autonomous weapons [114]. Critics, including the Electronic Frontier Foundation, argued that OpenAI's contractual "red lines" contained imprecise language that could permit surveillance activities. CEO Sam Altman admitted the rollout "was definitely rushed, and the optics don't look good" before the contract was amended on March 2, 2026, with strengthened protections [115].
The 2024-2025 restructuring from non-profit to for-profit drew sustained criticism. Opponents argued that OpenAI had received substantial tax benefits and donor contributions under the assumption it would prioritize AI's social benefits, not shareholder returns [39]. The attorneys general of California and Delaware negotiated conditions on the conversion, and the final structure (with the OpenAI Foundation retaining approximately 26% ownership and board appointment rights) represented a compromise [2]. SoftBank's $30 billion investment was contingent on the successful completion of this conversion [64].
March 2023: ChatGPT bug exposed users' chat histories and payment information [65]
2023-2025: Multiple investigations by European data protection authorities
Italy ban (March-April 2023): Temporary ban over GDPR concerns, lifted after compliance measures
Ongoing concerns about training on personal data without consent
| Company | Key Products | Strengths | Market Position |
|---|---|---|---|
| Google DeepMind | Gemini 2.5, Gemma | Search integration, compute resources, YouTube data | Consumer market surge (18-25% chatbot share by early 2026); leading in enterprise (21%) |
| Anthropic | Claude 4, Constitutional AI | Safety focus, long context, enterprise trust | Enterprise leader (~40% share); capturing 73% of first-time enterprise AI buyer spending (March 2026) [121] |
| Meta AI | Llama 4, SAM 2 | Open source, social integration, free models | Developer ecosystem dominance; largest open-weight model provider |
| DeepSeek | DeepSeek-R1, V3 | Extreme cost efficiency, open source, competitive reasoning | Market disruptor; trained R1 for ~$6M vs. hundreds of millions at OpenAI |
| xAI | Grok-3, Aurora | Real-time data, X integration | X/Twitter ecosystem; Grok overtaking Perplexity in usage |
| Mistral AI | Mistral Large 2, Mixtral | European, open models | EU market, sovereign AI deployments |
| Cohere | Command R+, Embed, Rerank | Enterprise focus, RAG | B2B, specialized applications |
| Stability AI | Stable Diffusion 3, StableVideo | Open source, community | Creative tools, local deployment |
OpenAI remains the overall market leader, but its dominance has narrowed significantly:
Consumer Chatbot Market Share: ChatGPT holds approximately 68% as of January 2026, down from 87% one year earlier. Google Gemini surged to 18% from 5.4%, while Anthropic's Claude holds roughly 2% [97]
Mobile App Market Share: ChatGPT's app share fell from 69% (January 2025) to 45% (early 2026), with Gemini growing to 25% [97]
Enterprise LLM Market: Anthropic has overtaken OpenAI, holding approximately 40% of enterprise market share vs. OpenAI's 27% and Google's 21% as of 2025. Among first-time enterprise AI buyers, Anthropic now captures over 73% of spending, a dramatic shift from 50/50 just weeks prior [98][121]
Consumer AI Users: ChatGPT still leads with 900M weekly users (February 2026)
Developer Tools: GitHub Copilot has ~70% of AI coding assistant market, but faces increasing competition from Cursor, Windsurf, and Devin
Revenue Leadership: Largest AI revenue among pure-play AI companies ($25B+ annualized run rate)
The narrowing of OpenAI's lead has been driven by several factors: Google's distribution advantages through Android and Search, Anthropic's strong reputation for reliability and safety in enterprise settings, Meta's free and open Llama models attracting developers, and DeepSeek's demonstration that competitive AI models can be built at a fraction of the cost [97][98]. OpenAI's response has included a strategic shift toward enterprise with the Frontier platform and Frontier Alliances, as well as plans to consolidate its consumer products into a single desktop superapp [117][121].
| Industry | Use Cases | Adoption Rate | Key Implementations |
|---|---|---|---|
| Healthcare | Diagnosis assistance, drug discovery, medical documentation | 67% of major hospitals | Mayo Clinic, Johns Hopkins |
| Finance | Risk analysis, fraud detection, algorithmic trading | 89% of investment banks | Goldman Sachs, JPMorgan |
| Education | Personalized tutoring, content creation, assessment | 45% of universities | Khan Academy, Duolingo |
| Legal | Contract analysis, legal research, document review | 72% of large firms | Casetext, Harvey AI |
| Software | Code generation, debugging, documentation | 78% of tech companies | Microsoft, Stripe |
| Creative | Content generation, design, music composition | 56% of media companies | Shutterstock, Adobe |
| Retail | Customer service, inventory, personalization | 61% of major retailers | Shopify, Instacart |
| Manufacturing | Quality control, predictive maintenance, supply chain | 42% adoption | Tesla, General Motors |
| Government & Defense | Administrative operations, cyber defense, data analysis | Expanding (2026) | U.S. Department of Defense, federal agencies via AWS GovCloud [114][118] |
Positive Contributions:
Democratized access to AI capabilities
Accelerated scientific research and drug discovery
Enhanced educational resources globally
Improved accessibility tools for disabilities
Boosted productivity across industries (75% of workers report improved speed or quality of output with AI at work) [50]
Challenges and Concerns:
Job displacement in certain sectors
Misinformation and deepfake risks
Academic integrity concerns
Digital divide and access inequality
Environmental impact of compute requirements
Military and surveillance applications of AI technology [114]
As of early 2026, OpenAI's trajectory is focused on agentic AI, infrastructure expansion, consumer hardware, enterprise platforms, and a potential public offering:
| Timeline | Planned Developments | Strategic Goals |
|---|---|---|
| 2026 | GPT-5.4+ model iterations, Codex expansion, Frontier platform scaling, potential IPO filing, io hardware product development, desktop superapp consolidation, Jobs Platform launch | Autonomous agent capabilities, public market access, consumer devices, enterprise dominance |
| 2027 | Multimodal native models, robotics integration | Physical world interaction |
| 2028 | Domain-specific expert systems | Professional replacement tools |
| 2029 | AGI achievement target, profitability | Mission accomplishment |
2026: $29-30 billion revenue, potential IPO consideration
2027: $50+ billion revenue, international expansion
2028: $75+ billion revenue, vertical integration
2029: $100 billion revenue target, break-even goal
Post-2029: Transition based on AGI achievement
OpenAI's projected annual cash burn is expected to rise significantly, from roughly $17 billion in 2026 to $35 billion in 2027 and $47 billion in 2028, making the liquidity of public markets a near-necessity for sustaining operations [8]. In February 2026, OpenAI reset its compute spending expectations, targeting approximately $600 billion in cumulative spending by 2030 [99].
1 million+ token context windows (achieved with GPT-5.4, March 2026)
Native computer-use capabilities for AI agents (achieved with GPT-5.4, March 2026)
Open-weight model release (achieved with gpt-oss, August 2025)
Enterprise AI agent platform (achieved with Frontier, February 2026)
AI-native web browser (achieved with Atlas, October 2025)
Real-time multimodal reasoning
Long-term memory and planning
Scientific discovery automation
Human-level problem solving across domains
The following dedicated pages cover major OpenAI releases and platform changes referenced throughout this article:
