OpenAI has changed the model most ChatGPT users meet first.
Rather than staging a major launch, the company has quietly replaced GPT-5.3 Instant with GPT-5.5 Instant as the default model inside the chat box used by hundreds of millions of people.
The most notable part is not the naming. It is the number OpenAI chose to publish: 52.5%.
The numbers are now public
OpenAI did not lean only on broad claims about better quality. In higher-risk categories such as medical, legal and financial questions, GPT-5.5 Instant produces 52.5% fewer hallucinatory statements than the previous generation, according to the company. In conversations where users had previously flagged factual mistakes, inaccurate statements were down 37.3%.
The benchmark scores moved higher as well:
| Test | GPT-5.3 Instant | GPT-5.5 Instant |
|---|---|---|
| AIME 2025 | 65.4 | 81.2 |
| MMMU-Pro | 69.2 | 76 |
AIME is a competition-level math benchmark, so a jump from 65.4 to 81.2 is material. MMMU-Pro tests doctoral-level scientific reasoning, and the new model moves into the 80% range.
Still, the Instant line has never been only about leaderboard scores. Its job is low latency plus reliability. When someone asks whether a medicine can be taken with ibrutinib, getting a wrong answer in three seconds is more dangerous than waiting thirty seconds for a correct one.
Fewer replies cluttered with emoji
OpenAI says the model's answers are more compact and more directly on point, without losing substance, while keeping the warmth and personality that make ChatGPT comfortable to use.
In practical terms, that means less filler.
OpenAI says word count is down by about 30% and line count by about 29%. It also says the model avoids elements that make replies look messy, including unnecessary emoji.
If ChatGPT's habit of sprinkling replies with symbols such as targets, rockets and sparkles has been wearing on users, this update is OpenAI's answer.
How the change rolls out
GPT-5.5 Instant immediately replaces GPT-5.3 Instant as ChatGPT's default model.
- Plus and Pro users: available on the web now, with mobile access rolling out.
- Free, Go, Business and Enterprise users: rollout is expected over the coming weeks.
- Developers: the API name is
chat-latest, while GPT-5.3 remains available to paid users for a three-month transition period.
There is also a quieter but useful privacy and control change: ChatGPT memory sources are now visible across all models. Users can see which conversation or file contributed a memory, then delete or correct it. If a chat is shared with someone else, those memory sources are not shown to the recipient.
Why OpenAI is publishing hallucination figures
For the past two years, major model companies have often talked about hallucinations in vague language, or avoided the subject entirely.
OpenAI's decision to put a percentage on the table is not hard to read. Too many ChatGPT users have been hurt by wrong answers in financial, medical and legal settings, and the public pressure has accumulated. The backlash around the retirement of GPT-4o in February, when some users described it as their best friend, also taught OpenAI that dependence on the default model runs deeper than product teams may expect.
A default-model change will not necessarily get a keynote or dominate headlines. But it shapes what hundreds of millions of people receive the next time they ask ChatGPT a question.
OpenAI has now made 52.5% the number to watch.
The next question is whether Claude and Gemini will follow by publishing their own hallucination rates.
Sources: CocoLoop, GPT-5.5 Instant: smarter, clearer, and more personalized (OpenAI official blog); OpenAI releases GPT-5.5 Instant, a new default model for ChatGPT (TechCrunch); OpenAI releases GPT-5.5 Instant update to make ChatGPT smarter with fewer emoji (9to5Mac)