What OpenAI is
OpenAI is an AI research company and product developer best known for creating ChatGPT and the GPT family of large language models. Founded with the goal of building artificial general intelligence (AGI) — AI that can perform any intellectual task a human can — it has become the most publicly visible organization in the modern AI boom. Its products are used by hundreds of millions of people, and its research has shaped how the entire industry thinks about building and scaling AI systems.
Why it matters
Before ChatGPT, powerful AI models existed mostly inside research labs and developer APIs. OpenAI's November 2022 launch of ChatGPT changed that: it gave anyone with a browser a way to have a real conversation with a capable AI, and the world noticed. Within months, AI had moved from a niche technical topic to front-page news. That moment didn't just grow OpenAI's business — it accelerated investment, competition, and public debate across the entire field.
The research foundation
OpenAI's influence starts with its research. A few papers in particular set the direction for the whole industry:
- GPT-1 (2018) showed that pre-training a model on lots of text, then fine-tuning it for specific tasks, was a powerful and general approach.
- Scaling Laws (2020) gave labs a mathematical roadmap: model performance improves predictably as you add more compute, data, and parameters. This paper justified the race to build ever-larger models.
- GPT-3 (2020) put that roadmap into practice with a 175-billion-parameter model that could perform a huge range of tasks with just a few examples — no task-specific training required.
- CLIP (2021) extended the same ideas to images, teaching a model to understand pictures using natural language descriptions.
These weren't just academic contributions. They became the blueprint that OpenAI — and its competitors — followed to build the products people use today.
The model family: from GPT-4 to GPT-5.5
OpenAI's model releases have accelerated sharply. GPT-4 (March 2023) was the first model to accept both images and text as input and to perform at a human level on a range of professional benchmarks. GPT-4o (May 2024) went further, handling audio, vision, and text in a single unified model in real time.
Then came a new idea: instead of just training models to be smarter, what if you let them think longer before answering? That's the core of the o1 series (September 2024), which uses a technique called chain-of-thought reasoning — the model works through a problem step by step, like showing its work — to tackle hard math, science, and coding problems. OpenAI called this "inference-time compute," a new axis of improvement beyond just making the model bigger.
The GPT-5 family, released from August 2025 onward, represents the current frontier: GPT-5, GPT-5.2, GPT-5.4, and GPT-5.5 have shipped in rapid succession, each claiming improvements in reasoning, coding, and speed. GPT-5.4 introduced a 1-million-token context window — meaning it can read and reason over roughly a book's worth of text in a single session — along with the ability to control a computer directly (called "computer use"). Smaller, faster variants like GPT-5.4 mini and nano target cost-sensitive applications and automated pipelines.
Doing new science
One of the most striking recent developments is OpenAI's models contributing to genuine scientific discovery — not just summarizing known facts, but producing new ones.
- GPT-5.2 proposed a novel formula for a gluon amplitude in theoretical physics, which was subsequently proved and verified by researchers.
- An OpenAI model disproved the Erdős planar unit distance conjecture, an 80-year-old open problem in discrete geometry — reportedly at a compute cost under $1,000.
These results suggest that frontier AI is beginning to function as a research tool capable of original work, not just a very fast search engine.
The business: funding, infrastructure, and partnerships
OpenAI has raised capital at a scale that would have seemed implausible a few years ago. In February 2026 alone, it announced a $110 billion investment round at a $730 billion valuation, with SoftBank, NVIDIA, and Amazon among the backers. A separate earlier round brought in $122 billion more. These funds are earmarked for compute infrastructure and model development.
The Stargate Project, announced in January 2025, is a joint venture targeting up to $500 billion in AI infrastructure investment in the United States — one of the largest such commitments in history.
On the cloud side, Microsoft has been OpenAI's primary partner since a $1 billion investment in 2019 that made Azure OpenAI's exclusive cloud provider. That relationship has since expanded enormously, but OpenAI has also struck a major deal with Amazon Web Services — a stateful runtime environment for AI agents on AWS, backed by a $15 billion Amazon investment and a $100 billion compute commitment — while keeping Microsoft's exclusive rights to host stateless API calls intact.
Government, defense, and the safety debate
OpenAI signed a formal contract with the U.S. Department of War in early 2026, allowing its models to be used "for all lawful purposes" in classified environments, with negotiated safety guardrails. This put it in direct contrast with Anthropic, which was formally designated a national security supply-chain risk after refusing to remove restrictions on its models for autonomous weapons and mass domestic surveillance.
The episode is a live example of the tension every frontier AI lab faces: how much to restrict what your models can do, and who gets to decide.
Open weights and the competitive landscape
For most of its history, OpenAI kept its most capable models proprietary. In August 2025, it released gpt-oss-120b and gpt-oss-20b under the Apache 2.0 license — open-weight models that anyone can download, run, and build on. This was a notable strategic shift, likely in response to strong open-source competition (including DeepSeek's R1, which claimed performance parity with OpenAI's o1 at a fraction of the cost).
Where it's heading
The pattern across these events is one of simultaneous expansion on every front: more capable models released faster, more infrastructure to train and serve them, more partnerships to distribute them, and more specialized versions (like GPT-Rosalind for life sciences) targeting specific industries. The open-weights move suggests OpenAI is also competing for developer mindshare, not just enterprise contracts. The core tension — building the most powerful AI in the world while managing what it can be used for — shows no sign of resolving soon.
