OpenAI rolls out GPT-5.5 Instant as ChatGPT’s new default, touting fewer high-stakes “hallucinations”

OpenAI is swapping out the engine that powers ChatGPT for most people, and it says the new default model makes fewer things up when the stakes are high.

The company is rolling out “GPT-5.5 Instant” to all ChatGPT users, replacing “GPT-5.3 Instant.” OpenAI claims the upgrade delivers clearer, more direct answers and cuts hallucinated claims by 52.5% on sensitive topics like medicine, law, and finance, areas where a confident-sounding mistake can do real damage.

Alongside the model change, OpenAI is also pushing a new transparency feature called “memory sources,” designed to show users what personal context ChatGPT relied on, like prior chats, uploaded files, or connected services such as Gmail, when it tailors a response.

GPT-5.5 Instant becomes the default for everyone

For most users, the change will be automatic. GPT-5.5 Instant is becoming the standard model that answers when you open ChatGPT and type a question, no settings tweaks required.

OpenAI positions the “Instant” line as optimized for speed and efficiency, which is why it’s the default. The practical implication: the version most people actually use day-to-day is the one OpenAI is now betting will be more reliable.

Paid users will still be able to select GPT-5.3 Instant for about three months through model settings before it’s removed. For everyone else, the rollout is gradual, which can make the experience feel inconsistent, like ChatGPT is “different” from one day to the next.

OpenAI is also aligning its developer offering around the same idea. The company says GPT-5.5 Instant is available via its API under the name “chat-latest,” signaling a push toward a more consistent “default” model across both the consumer app and developer tools.

OpenAI claims a 52.5% drop in hallucinations on medicine, law, and finance

The headline number is OpenAI’s claim that GPT-5.5 Instant produces 52.5% fewer hallucinated “claims” than GPT-5.3 on “high-stakes” topics, specifically medicine, legal questions, and financial guidance.

OpenAI also says the model reduces inaccurate statements by 37.3% in especially difficult conversations that users had previously flagged for factual errors, essentially, scenarios where the system had already been caught getting things wrong and was tested again.

If those improvements hold up in real-world use, they could matter most in the kinds of prompts people increasingly throw at chatbots: drug interaction questions, explanations of legal processes, or breakdowns of financial ratios. But OpenAI’s numbers come from internal testing, not independent audits, and performance can vary depending on language, topic, and how a question is asked.

The bottom line for users doesn’t change: ChatGPT can speed up research and drafting, but it’s still not a doctor, lawyer, or financial advisor, and it still needs verification when decisions and money are on the line.

“Memory sources” aims to show what personal context ChatGPT used

The other major shift is about transparency, and privacy. OpenAI is rolling out “memory sources” across ChatGPT models, letting users see what context the system used to personalize an answer.

That context can include past conversations, files you’ve shared, and, if you’ve enabled it, data from connected services like Gmail. On the web, users can access the feature via a “Sources” icon under a response.

The pitch is simple: instead of guessing whether ChatGPT is “remembering” something you told it months ago, or just inventing a detail, you can see what it pulled from. And if the context is outdated or wrong, OpenAI says users can correct or delete it.

For everyday use, this could reduce a common frustration: the bot confidently leaning on old assumptions, your old city, old dietary preferences, old work constraints, and steering recommendations in the wrong direction. But it also makes the personalization tradeoff more visible, which could unsettle users who didn’t realize how much context might be in play.

Advanced personalization rolls out first to paying users on the web

OpenAI is drawing a clear line between access to the new default model and access to the most powerful personalization features.

GPT-5.5 Instant is rolling out to free users, too. But OpenAI says advanced personalization, especially richer use of context from chats, files, and connected Gmail, will arrive first for Plus and Pro subscribers on the web.

The company says the feature set will expand to mobile “soon,” and then to other tiers, including Free, Go, Business, and Enterprise, over the coming weeks. For workplaces, the memory-and-sources system could raise new questions about compliance and internal data policies, even if OpenAI provides user controls.

Shorter answers, fewer emojis, and a “Thinking” mode for harder tasks

OpenAI is also promising a tone shift: tighter, more direct responses with less unnecessary formatting, fewer emojis, and fewer follow-up questions that don’t add value.

That might sound cosmetic, but it’s a real quality-of-life change for heavy users, especially on phones, where long, hedged answers can turn into endless scrolling.

OpenAI also says GPT-5.5 Instant improves image analysis, performs better on STEM questions, and makes smarter calls about when to use web browsing (which can slow responses down). For more complex tasks, ChatGPT can also switch to “GPT-5.5 Thinking,” a deeper-reasoning mode that can be selected manually.

The bigger bet behind all of it: most people never touch model settings. If OpenAI wants fewer factual blowups in the real world, it has to improve the default experience, right where the majority of users live.