On May 5, 2026, OpenAI deployed GPT-5.5 Instant as ChatGPT's new default model for all users, with 52.5% fewer hallucinations on high-stakes queries.
A silent but structuring replacement
Exactly two months after the introduction of GPT-5.3 Instant on March 3, 2026, OpenAI carried out another rotation of its base model. This bi-monthly rhythm is starting to look like an established cadence: the company adjusts its consumer engine with clockwork regularity, without any thunderous announcements. The stated goal this time is not to impress with unprecedented reasoning capabilities, but to make ChatGPT more reliable on a daily basis for the hundreds of millions of people who use it every week.
OpenAI presented this update in a post published on its official website on May 5, 2026. The message is clear: the Instant model is the "everyday engine," the one that the vast majority of ChatGPT users interact with at every turn. Improving this specific link, even modestly, has a multiplier effect.
The fight against hallucinations, in numbers
The main focus of this update is the reduction of factual errors. According to OpenAI's internal evaluations published in its official post, GPT-5.5 Instant produced 52.5% fewer hallucinatory statements than GPT-5.3 Instant on high-stakes queries covering medicine, law, and finance. For conversations that users had flagged as particularly prone to errors, the reduction reaches 37.3%.
These improvements are confirmed on external benchmarks. According to data reported by TechCrunch, the model scores 81.2 on the AIME 2025 math test, compared to 65.4 for its predecessor. On the multimodal reasoning benchmark MMMU-Pro, GPT-5.5 Instant reaches 76 compared to 69.2 for GPT-5.3. OpenAI also highlights notable gains in image analysis, scientific reasoning (STEM), and web search management to supplement a response.
However, a nuance is necessary. These figures come from internal evaluations, and the company itself acknowledges that a model that hallucinates less can paradoxically encourage users to trust it more, even when it is still wrong. The reduction in hallucinations is real, but it does not mean their disappearance.
Less verbosity, fewer emojis
The other aspect of this update concerns style. According to data published by OpenAI, GPT-5.5 Instant uses 30.2% fewer words and 29.2% fewer lines compared to its predecessor to convey the same information. Responses are more direct, less cluttered with unnecessary follow-up questions, artificial formatting, and emojis.
OpenAI describes this evolution as seeking a balance: preserving the "warmth and personality" of ChatGPT while eliminating what the company calls "over-formatting." For an informal advice query, the model will no longer adopt the structure of a professional report when the user simply expected a practical and quick answer. This shift is not insignificant. When GPT-4o was withdrawn in February 2026, users launched petitions to prevent its deprecation, describing the model as a "friend" or "mirror." OpenAI is clearly calibrating its personality strategy with caution.
Deeper contextual memory
The third dimension of this update concerns personalization. GPT-5.5 Instant more effectively utilizes available context: past conversations, shared files, and even the content of a connected Gmail account if the user has authorized it. The goal, as stated by OpenAI in its official blog post, is for users not to have to repeat themselves from one session to another.
The most concrete novelty is the introduction of "memory sources" (memory sources), an interface available on all ChatGPT models. It allows users to see what context was used to personalize a response, to correct it, or to delete it. If a user shares a conversation with someone else, the memory sources are not transmitted. OpenAI emphasizes this point: the user remains in control of what the model knows about them, with the ability to delete entire conversations from memory or manually edit saved items.
This feature brings ChatGPT closer to a personal assistant in the full sense of the term, capable of adapting to your professional context, your habits, and your preferences without you having to rephrase your preferences each session.
Deployment: who can access it, and under what conditions
The deployment of GPT-5.5 Instant began on May 5, 2026, for all ChatGPT users, including free accounts. Advanced personalization with access to files and Gmail is being rolled out primarily for Plus and Pro subscribers on the web, before being extended to mobile apps and then to free users in the coming weeks. Memory sources are available for all consumer accounts on the web now, and then on mobile.
For paying users, GPT-5.3 Instant remains accessible for three months before being permanently removed. Developers access the new model via the API under the alias chat-latest.
We also need to place GPT-5.5 Instant within OpenAI's broader offering architecture. The previous month, the company had launched GPT-5.5 Thinking and GPT-5.5 Pro, presented as models suited for complex tasks: advanced coding, in-depth research, long workflows. GPT-5.5 Instant is the model for speed and everyday use. It is faster, designed for common interactions, and according to OpenAI, maintains the same latency as GPT-5.3.
This update confirms a fundamental trend in OpenAI's strategy: refining the experience for the vast majority of users rather than focusing efforts on cutting-edge performance reserved for a minority. For any user who opens ChatGPT several times a day, the question is not whether GPT-5.5 Instant can solve an Olympic-level math problem, but whether its answer to a medical or legal question will be truly more reliable. On this criterion, the figures published by OpenAI indicate significant progress.
No Comment! Be the first one.