ChatGPT Passes 1 Billion Users — and Gets a Memory That Decides for Itself
According to market data from Sensor Tower, ChatGPT passed one billion monthly active app users in May 2026 — roughly three years after launch, and faster than any previously measured app. Almost simultaneously, on June 4, 2026, OpenAI rolled out its biggest memory upgrade since the original ChatGPT launch: the „Dreaming V3” architecture, reaching US Plus and Pro subscribers first. The pairing is notable because it combines billion-scale reach with a memory that builds user profiles largely on its own. That very autonomy brings the system within scope of the EU AI Act transparency rules, which apply from August 2, 2026.
What actually changed
- ChatGPT reached around 1 billion monthly active app users in May 2026 per Sensor Tower (reported by Reuters on June 2, 2026) — a third-party analytics figure, not an official OpenAI number.
- The app grew 62 percent year over year by these figures; for context, apps like YouTube and TikTok took five to eight years to hit the same threshold.
- OpenAI launched „Dreaming V3” on June 4, 2026, first for Plus and Pro users in the United States; rollout to Free and Go users and further countries „over the coming weeks”.
- The system learns automatically from all past conversations, per OpenAI; the compute needed to serve it to free users was cut by roughly 5x.
- An arXiv study (February 2026, 2,050 memory entries from 80 users) found: 96 percent of memories were created by the system itself, only 4 percent on explicit user instruction.
What used to be the case
Memory in ChatGPT had been a comparatively simple feature: the system remembered individual facts a user explicitly asked it to store, or that it inferred from a conversation. Users could review and delete stored entries, but the mechanism stayed manageable — a notepad, not a profile.
Regulation was looser, too. The EU AI Act had entered into force in 2024, but its Article 50 transparency obligations only took effect later. Until then, disclosing what a system stores about a user and how it draws inferences was largely a matter of the provider’s terms of service — not an enforceable legal duty.
What applies now
1. Billion-scale reach — read with caution. The one-billion figure refers to the mobile app and comes from analytics provider Sensor Tower, not an OpenAI announcement. It evidences the broad global adoption of AI assistants, but it is a third-party estimate. A competing count put the user base at around 1.1 billion while market share declined — the absolute number and its interpretation vary by source.
2. Memory that builds profiles by itself. Dreaming V3 is no longer a notepad. Per OpenAI, the system learns automatically from the entire conversation history and weighs what it infers about a user. That shifts the function from „explicitly stored facts” to „continuously updated behavioral profile” — a qualitative leap, not just a bigger store.
3. The research record urges caution. The arXiv study „The Algorithmic Self-Portrait” analyzed 2,050 memory entries from 80 real users. Findings: 96 percent of entries were created without an explicit user request, 28 percent contained personal data protected under GDPR, 52 percent psychological assessments of the person. The sample is small and predates the Dreaming V3 rollout — as a directional signal on the mechanics of self-created profiles, it remains relevant.
Why it matters
The real news value lies not in the round user count but in the simultaneity. An assistant with billion-scale reach gets a memory that builds behavioral profiles largely without explicit consent — and that just weeks before the AI Act’s transparency obligations take effect in the EU. Building a behavioral profile that the user did not actively initiate sits squarely in the tension Article 50 addresses: people should know what they are interacting with and what is being processed about them.
For a privacy-first, EU-oriented architecture this is more than a footnote. The arXiv figure of 96 percent system-created memories captures the core tension: convenience (the model „knows” the user) against data sovereignty (the user controls what is stored). Both numbers — the user billion and the 96 percent — should be flagged cleanly as reporting and study figures, not as confirmed vendor facts. OpenAI has not, in any statement cited here, responded to the study’s numbers.
The efficiency angle matters too: only the roughly fivefold reduction in compute makes it feasible to serve such a memory to the free mass base at all. Profile-building thus scales toward the billion mark precisely when it is under the closest regulatory scrutiny.
What you can do now
Actively review your ChatGPT memory settings. Do not assume the system only stores what you approve. Look through the saved memories, delete what does not belong, and turn off automatic capture when discussing sensitive or business topics.
Treat AI memory as a data-protection topic in client and compliance conversations. From August 2, 2026, the AI Act’s transparency obligations apply — and not only to high-risk systems. Anyone embedding generative AI in products should clarify now what is stored about and inferred from users.
Separate reach PR from solid facts. The one-billion headline is a third-party estimate for the app, not the full product. Use it as an adoption signal, not an exact accounting figure.
Entdecke mehr
Headless without an API bill — how do you reach the best AI models for automation in 2026?
Provider comparison mid-2026: who has a headless mode, whose subscription still covers it — and why BYOK is the most stable foundation.
Glossar$/MTok (Cost per Million Tokens)
Standard pricing unit for AI APIs — cost in US dollars per one million processed tokens. Listed separately for input, output and sometimes cache.
LexikonFunction Calling / Tool Use
How an LLM uses tools: define a tool as a schema, the model picks the function and arguments, the result returns to the chat — the basis of every agent.