OpenAI on Monday released GPT-6, the next generation of its flagship model, with a feature the company is calling persistent context: a permanent, opt-in memory layer that travels with users across sessions, devices, and applications.
In a 40-minute keynote streamed from San Francisco, chief executive Sam Altman framed the launch as the company's first "post-chatbot" product. "The model now knows you," he said. "Not just what you said five minutes ago — what you've been working on for months."
What's actually new
Benchmark scores tell only part of the story. GPT-6 posts a 94.2% on the GPQA-diamond reasoning suite, narrowly edging out Anthropic's Claude Opus 5.1, but the real differentiator is engineering: a hybrid retrieval-augmented architecture that lets the model reach into a structured user history without re-prompting on every turn.
Early developer reports describe a meaningful drop in token spend on long-running workflows — agents that previously needed to re-read project context on every invocation can now skip that work entirely.
Regulatory questions
The European Commission's AI Office issued a same-day request for technical documentation under the AI Act's general-purpose model provisions. A spokesperson confirmed that "persistent personal memory" was being treated as a distinct risk category for the purposes of disclosure obligations.
OpenAI says memory is fully user-controlled and exportable. Industry observers will be watching whether the controls are clear enough to satisfy GDPR's right-to-erasure framework.