From One Paper: The Parallel Rise of OpenAI and Anthropic
From One Paper: The Parallel Rise of OpenAI and Anthropic
The paper was eight pages long. It had eight authors, all of them affiliated with Google Brain or Google Research. It was submitted to a conference on neural information processing systems in 2017, accepted, and presented. Its title was Attention Is All You Need. It described a new neural network architecture called the Transformer — a new way of teaching machines to process and generate language. Within the field, it was recognized quickly as significant work. Outside the field, it attracted almost no attention.
That paper is the origin point of everything that follows. Every major language model released in the years since — GPT-2, GPT-3, GPT-4, Claude, Gemini, LLaMA — is built on the Transformer architecture. The two companies at the center of this story, OpenAI and Anthropic, both owe their core technology to work done at Google in 2017. OpenAI existed when the paper was published, but as a nonprofit. Anthropic would not exist for another four years.
OpenAI was incorporated in December 2015 as a nonprofit artificial intelligence research laboratory. Its stated mission was to ensure that artificial general intelligence — AI capable of performing any intellectual task a human can — would benefit humanity broadly rather than accrue to any single entity. It was funded by a group of technology figures who collectively committed approximately $1 billion, though only a fraction of that was initially deployed.
The organization began hiring researchers. It published work on reinforcement learning, robotics, and game-playing AI. Its early releases included systems that learned to play video games at superhuman levels and a system called OpenAI Five that competed against professional players of the strategy game Dota 2. These were significant technical achievements. They were not language models, and they did not have obvious consumer applications.
In 2017, the Transformer paper appeared. OpenAI, like every other AI research organization, took note.
In June 2018, OpenAI published a paper describing GPT-1 — Generative Pre-trained Transformer. The model was trained on a large corpus of text and could generate coherent language across a range of tasks. It was a proof of concept: a demonstration that the Transformer architecture, scaled up and trained on enough data, could produce general-purpose language capabilities without being explicitly programmed for each task.
GPT-2 followed in February 2019. It was significantly larger than its predecessor — 1.5 billion parameters compared to GPT-1's 117 million — and its outputs were noticeably more coherent and fluent. OpenAI made an unusual decision at release: it declined to publish the full model, citing concerns that it could be used to generate large volumes of convincing misinformation. It released smaller versions instead, with the full model following later in the year after the organization concluded the risks had been overstated.
The decision attracted press coverage of a kind OpenAI had not previously received — coverage focused not on technical capability but on potential harm. It was the first time a language model release had generated that kind of public discussion. It would not be the last.
GPT-2's outputs were noticeably more coherent and fluent. OpenAI declined to publish the full model, citing concerns it could generate convincing misinformation at scale.
In March 2019, OpenAI announced a structural change. It created a new for-profit subsidiary — OpenAI LP — while retaining a nonprofit parent, OpenAI Inc. The structure was described as a "capped-profit" model: investors and employees could earn returns on their investment, but those returns were capped at one hundred times the initial investment. Any value beyond that cap would flow to the nonprofit mission rather than to shareholders.
The rationale was practical: training large language models required computing resources that a nonprofit could not easily fund through donations alone. The new structure would allow OpenAI to raise investment capital. It also meant OpenAI was no longer purely a research organization. It was, structurally, a company.
In May 2020, OpenAI published a paper describing GPT-3, a language model with 175 billion parameters — more than one hundred times larger than GPT-2. The scale produced capabilities that were qualitatively different from anything researchers had seen before. GPT-3 could write coherent essays, answer factual questions, generate functioning code, translate languages, and complete tasks it had never been explicitly trained for, simply by being shown a few examples in the prompt.
In June 2020, OpenAI made GPT-3 available to developers through a private API, with a waitlist for access. The response from the developer community was immediate and enthusiastic. Within weeks, people were building applications — writing tools, customer service bots, code assistants, tutoring systems — on top of a model they could access through a simple API call without understanding how it worked internally.
In July 2020, Microsoft announced a $1 billion investment in OpenAI and an exclusive licensing agreement covering GPT-3. Under the agreement, Microsoft gained the right to incorporate GPT-3's underlying technology into its own products and services. OpenAI gained access to Microsoft's Azure computing infrastructure and the capital to continue scaling its research.
It was OpenAI's first major commercial partnership. It would not remain its only one.
In the spring of 2021, a group of researchers left OpenAI. Among them were Dario Amodei, who had been OpenAI's VP of Research, and Daniela Amodei, who had been VP of Operations. They were joined by seven other former OpenAI employees. In total, the founding team numbered eleven people.
In May 2021, they incorporated a new company: Anthropic. It was structured as a public benefit corporation — a legal designation that requires the company to consider social impact alongside profit. Its stated focus was AI safety research: the study of how to build AI systems that behave reliably, that do not produce harmful outputs, and that remain aligned with human intentions as they become more capable.
Anthropic raised $124 million in a seed round from a group of investors that included former Google CEO Eric Schmidt and venture firm Spark Capital. The company began building its own large language model from the ground up.
OpenAI and Anthropic now occupied the same space — large language model research and development — with largely overlapping technical expertise and a shared starting point in the Transformer architecture. They would operate as separate companies, pursuing different approaches, from that point forward.
In March 2022, OpenAI released GPT-3.5 — a refined version of GPT-3, trained using a technique called reinforcement learning from human feedback, or RLHF. The technique involved human trainers rating model outputs and using those ratings to fine-tune the model's behavior. The result was a model that was not just more capable but more cooperative — better at following instructions, less likely to produce irrelevant or harmful responses.
Also in 2022, Anthropic published foundational safety research. One paper introduced a training approach called Constitutional AI, in which a model is given a written set of principles and trained to evaluate and revise its own outputs against those principles. The approach was designed to reduce reliance on extensive human feedback during the safety training process.
On November 30, 2022, OpenAI released ChatGPT. It was built on GPT-3.5 and wrapped in a conversational interface: a simple text box where a user could type a message and receive a response. It was described as a research preview. The expectation was modest early-adopter interest.
ChatGPT reached one million users in five days. It reached one hundred million users in two months. The interface — a plain text box requiring no technical knowledge — put a capable language model in front of anyone with an internet connection. The product demonstrated, at scale, that there was enormous latent demand for this kind of tool.
The AI era, in the public sense, began on November 30, 2022.
In January 2023, Microsoft announced a new multiyear investment in OpenAI reported at $10 billion. The deal deepened the existing partnership: OpenAI's models would power Microsoft's products across Bing, Office, and Azure. Microsoft's computing infrastructure would underpin OpenAI's research and deployment.
In February 2023, Google released Bard, a conversational AI built on its own large language model, LaMDA. A promotional video showed Bard answering an astronomy question incorrectly. Alphabet's stock fell approximately 7 percent the day after the advertisement aired. Bard's public reception was cautious. The episode illustrated how quickly reputational risk could materialize in a high-visibility product category.
In March 2023, OpenAI released GPT-4. It outperformed GPT-3.5 on standardized professional and academic benchmarks — bar exams, medical licensing tests, graduate admissions exams — by significant margins. It was also multimodal, capable of processing both text and images as inputs. It became the engine behind the premium tier of ChatGPT.
Also in March 2023, Anthropic released Claude — its first publicly available language model. Claude was made available through a limited research preview and through an API for developers. Anthropic described Claude's design as incorporating Constitutional AI principles: the model had been trained to be helpful, harmless, and honest, with safety considerations embedded in the training process rather than applied as a filter on top of it.
By mid-2023, the field had expanded beyond OpenAI and Anthropic. Meta released LLaMA, a family of open-source language models that researchers and developers could run on their own hardware without API access or usage fees. Dozens of companies began building products on top of the leading models. The competitive landscape that had been effectively a one-company story for most of 2022 had become, within months, a crowded field.
In March 2024, Anthropic released the Claude 3 family: three models — Haiku, Sonnet, and Opus — calibrated for different combinations of speed, cost, and capability. Claude 3 Opus, the most capable of the three, outperformed GPT-4 on several benchmark categories at release, including graduate-level reasoning, multilingual mathematics, and coding. It was the first time an Anthropic model had led the field on widely-used capability benchmarks.
OpenAI responded in May 2024 with GPT-4o — a new version of GPT-4 with significantly improved speed, lower cost, and native multimodal capabilities including real-time voice interaction. The release was accompanied by a live demonstration that showed the model conducting natural spoken conversations, expressing something resembling affect in its tone, and responding to visual inputs in real time.
Google rebranded Bard as Gemini in February 2024 and released a series of Gemini models across multiple capability tiers. The benchmark competition among the three leading labs — OpenAI, Anthropic, and Google — became a regular feature of the technology press cycle, with each new release accompanied by detailed comparisons across standardized tests.
In late 2024, a Chinese AI laboratory called DeepSeek released a series of models that matched or exceeded the performance of leading Western models on several benchmarks, at a reported training cost dramatically lower than comparable American models. The release was widely noted in the industry. It demonstrated that the leading position in language model capability was not stable and that significant advances could originate outside the small group of well-funded American labs.
DeepSeek demonstrated that the leading position in language model capability was not stable, and that significant advances could originate outside the small group of well-funded American labs.
Through 2025, the competitive dynamics between OpenAI and Anthropic shifted in a direction that the consumer headlines had not fully anticipated. Enterprise adoption — large organizations integrating AI into internal workflows, customer-facing products, and professional services — became the primary growth vector for both companies.
An industry index tracking AI tool usage across thousands of companies showed Anthropic surpassing OpenAI in enterprise adoption during 2025. The data point reflected a pattern that had been building for several quarters: large organizations with legal, compliance, and reputational concerns about AI outputs tended to favor Anthropic's models, citing their safety training and their performance on tasks requiring extended reasoning and careful instruction-following.
OpenAI retained a significant lead in consumer mindshare — ChatGPT remained the most widely recognized AI product globally — and in total API usage volume across all customer categories.
Both companies expanded their model lineups. Anthropic released Claude 3.5 Sonnet in mid-2025, followed by Claude 3.5 Haiku and an updated Claude 3.5 Opus. OpenAI released the o1 and o3 model families, which incorporated extended reasoning capabilities — models that generated explicit reasoning chains before producing final outputs, improving performance on complex multi-step problems. The capability distinction between models designed for speed and low cost versus models designed for extended reasoning became a standard feature of how both companies positioned their product lines.
Both companies also moved aggressively into what the industry called agentic AI — systems capable of taking sequences of actions over time rather than simply responding to individual prompts. Anthropic released Claude Code, a command-line tool for software development that could read and modify codebases, run tests, and propose fixes across complex projects. OpenAI released Operator, a system designed to complete tasks in web browsers without human guidance at each step.
In early 2025, OpenAI closed a $40 billion funding round that valued the company at $300 billion — the largest private technology funding round recorded to that point. Anthropic raised $3.5 billion in a separate round that valued the company at approximately $61 billion. Both companies were, by any measure, among the most valuable private companies in the world. Neither position would hold for long.
In 2026, both companies moved toward the public markets in the same month. OpenAI completed its conversion from a capped-profit structure to a public benefit corporation, with the nonprofit parent retaining a 26 percent equity stake. On May 22, 2026, OpenAI filed a confidential draft registration statement with the Securities and Exchange Commission, targeting a public listing as early as September 2026. Goldman Sachs and Morgan Stanley are leading the offering.
On May 28, 2026, Anthropic closed a $65 billion Series H funding round at a post-money valuation of $965 billion — its largest raise and, by reported valuation, its last private financing before a public offering. The round was co-led by Altimeter Capital, Dragoneer, Greenoaks, Sequoia Capital, Capital Group, Coatue, and D1 Capital Partners, with participation from institutional investors including Blackstone, Brookfield, Fidelity, and strategic partners Samsung, SK Hynix, and Micron.
The competitive picture in mid-2026 reflected eight years of parallel development from a shared starting point. OpenAI had the most widely recognized consumer product, the deepest integration into Microsoft's enterprise software ecosystem, and a last reported private valuation of $852 billion. Anthropic had the strongest position in regulated enterprise categories, a reputation for safety-focused development that had attracted large commitments from Amazon and Google, and a post-Series H valuation of $965 billion. Both companies were targeting public listings for the latter half of 2026.
The Transformer architecture that eight Google researchers published in 2017 remained the foundation of every major language model either company had produced. Neither OpenAI nor Anthropic had originated the architecture. But both companies had built, on top of it, organizations of several thousand employees, products used by hundreds of millions of people, and businesses valued in the tens of billions of dollars.
Tech Reader Magazine covers the ideas, institutions, and technologies reshaping the world — with the depth and editorial independence that daily news cycles rarely allow. Each piece examines not just what happened, but what it means: for the industry, for the organizations navigating it, and for the broader relationship between technology and human judgment.