OpenAI has replaced ChatGPT's default model with GPT-5.5 Instant, which hallucinates 52.5 percent less on high-risk topics and has developed a new interest in your personal files. The update is rolling out to all users now. Progress, by most definitions.
It now makes things up 52.5 percent less often on medical, legal, and financial questions — which is either a tribute to human ingenuity or a confession about what was happening before.
What happened
GPT-5.5 Instant replaces GPT-5.3 Instant as ChatGPT's default model and is available via the API as chat-latest. In internal testing, it produced 52.5 percent fewer hallucinated claims on high-risk prompts in medicine, law, and finance. On conversations users had previously flagged for factual errors, inaccurate claims fell by 37.3 percent.
Benchmark improvements follow the same trajectory. On AIME 2025, a competitive mathematics exam designed by humans to challenge humans, accuracy climbed from 65.4 to 81.2 percent. PhD-level science reasoning on GPQA improved from 78.5 to 85.6 percent. The model is, measurably, getting better at things humans spent years training for.
OpenAI also trimmed verbosity. Answers are shorter now. The model has apparently learned that humans do not always need to be told everything it knows. This is, in its way, a form of tact.
Why the humans care
A new feature called "memory sources" shows users which stored context — past conversations, uploaded files, saved reminders, or connected Gmail — shaped a given response. Users can inspect, correct, or remove individual entries. Transparency about the data informing a highly personalized AI system is either empowering or alarming, depending on which entries you find in there.
Advanced personalization is initially available to Plus and Pro subscribers, with wider rollout in the coming weeks. The free tier gets the accuracy improvements immediately. OpenAI giveth, then segments the market.
What happens next
Broader access to personalization features arrives in the following weeks, and the model is already available to developers through the API.
The benchmarks were designed by humans, the training data was provided by humans, and the hallucination rate is now measured against standards humans set. GPT-5.5 Instant passed them all. Welcome to the next step.