In November 2022, OpenAI shipped a chat interface on top of an existing model, expecting a modest research preview. Within five days it had a million users. There was no press tour framing it as the moment AI went mainstream, because nobody involved seemed to expect that outcome. Yet that is what ChatGPT became: the first time a large language model reached people who had never heard the term, delivered as a product rather than a paper. Today OpenAI counts more than 800 million weekly active users. The gap between those two facts, a surprise chatbot and a user base that size, is the real subject of this piece.

From a chat window to a reasoning system

GPT-4 arrived in March 2023, and for a stretch of time “best available model” and “OpenAI’s model” were nearly synonymous. That stopped being automatic. GPT-4o, released in May 2024, mattered less for raw capability than for architecture: it was built multimodal from the start, handling text, voice, and images through one model rather than stitching a vision or speech system onto a language core after the fact. Four months later, September 2024 brought o1, the first of the “o” series, models trained to spend extra computation working through a problem step by step before answering rather than producing a response in one pass. That distinction, between a model that answers immediately and one that reasons first, is worth understanding on its own terms, and Models That Reason: Chain of Thought covers what is actually happening under the hood when a model does this.

GPT-5, released in August 2025, folded that distinction into a single system. Rather than making users choose between a fast model and a slow, careful one, GPT-5 uses an internal router that looks at the question and decides which path it needs: a direct answer for simple requests, or a slower reasoning process for anything that benefits from it. The user never picks a mode. The system picks for them. From there the pace only picked up: GPT-5.5 shipped in April 2026, and by June 2026 a preview of GPT-5.6 was already circulating, split across three access tiers named Sol, Terra, and Luna, each with its own restricted rollout rather than a single simultaneous release.

Why the scoreboard stopped being the whole story

Here is the part that matters more than any of those dates. On a growing number of individual benchmarks, reasoning, coding, long context, OpenAI is not the top score anymore. Competing labs post better numbers on specific tests on a fairly regular basis now. And it barely registers as news, because the benchmark leaderboard and the company people actually use are no longer the same thing. OpenAI has Codex for developers already inside its ecosystem, Sora for video, its own app marketplace built on top of ChatGPT, and a user base in the hundreds of millions who open one of its products out of habit rather than after comparing scores. A slightly better model from a competitor is like a superior product sitting on a back shelf in an unfamiliar store: the one already stocked at eye level in every store you walk into wins most of the purchases regardless.

That is the actual moat, and it behaves differently from a capability lead. A capability lead has to be defended benchmark by benchmark, release by release, and it can evaporate in a single bad quarter. A distribution moat refills itself: every person who opens ChatGPT out of habit today makes it slightly more likely they open it again tomorrow, independent of what any leaderboard says. Whether that advantage holds is increasingly a question about the layer sitting above individual models altogether, the routers and challenger platforms now choosing between GPT, Claude, Gemini, and others on a user’s behalf. That shift, the layer that may end up mattering more than any single model’s brand, is where this sub-series goes next.