On May 5, 2026, OpenAI announced the update of ChatGPT's default model to GPT-5.5 Instant, replacing GPT-5.3 Instant. The new version produces 52.5% fewer hallucinated claims on high-stakes prompts and introduces an improved personalization system based on conversation history and connected accounts.
Key highlights
- GPT-5.5 Instant is the new default ChatGPT model for all users, available from May 5, 2026
- 52.5% fewer hallucinations on prompts in medicine, law, and finance compared to GPT-5.3 Instant
- 37.3% fewer inaccurate claims in conversations flagged by users as factually incorrect
- New "memory sources" feature shows what data from past conversations was used to personalize responses
- GPT-5.3 Instant remains available to paid users for three months via model configuration settings
A change at the default position
ChatGPT's default model is the product used by hundreds of millions of people every day. Any update at this layer has a direct global-scale effect — even when changes are evolutionary rather than breakthrough. GPT-5.5 Instant is available to all ChatGPT users from May 5, 2026, and simultaneously as `chat-latest` in the API.
GPT-5.3 Instant will remain available to paid users for three months via model configuration settings before being retired. In the API, `chat-latest` now automatically points to GPT-5.5 Instant — which means potential breaking changes for developers building on this endpoint without pinning a specific version.
Accuracy: the numbers and their context
The primary measurable change is hallucination reduction. In OpenAI's internal evaluations, GPT-5.5 Instant produced 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts covering medicine, law, and finance. In conversations flagged by users as factually incorrect, the number of inaccurate claims fell by 37.3%.
The methodological context matters: these are the results of OpenAI's own internal evaluations, not independent benchmarks. OpenAI describes GPT-5.5 Instant as a "generally smarter and more capable" model for everyday tasks — a characterization that is inherently difficult to verify externally with precision.
The product documentation presents a worked example involving an algebra problem: GPT-5.3 Instant failed to detect a student's algebraic mistake and incorrectly concluded there was no real solution. GPT-5.5 Instant initially also confirmed the incorrect solution, but then independently detected the inconsistency when substituting, identified the actual algebraic error, and computed the correct solution using the quadratic formula.
Personalization with user control
The second axis of the update is personalization. GPT-5.5 Instant is more effective at drawing on context from past conversations, uploaded files, and — for users who have connected it — Gmail. According to OpenAI, the model is faster at searching conversation history and more intelligently decides when personalization genuinely improves a response.
The key new feature is "memory sources" — a context indicator. When a response is personalized, the user can see which data was used: saved memories or specific previous conversations. These can be deleted or corrected if they are outdated. Memory sources are not visible to others when a conversation is shared.
OpenAI explicitly acknowledges a limitation: the memory sources view may not show all factors that shaped a response — for example, it may display only the most relevant past conversations, not all those the model searched. The company says it will continue improving this view in future updates.
Less text, better communication
The update also includes changes to text generation style. GPT-5.5 Instant is designed to produce responses that are "tighter without losing substance" — free of excessive formatting, unnecessary follow-up questions, and gratuitous emojis. OpenAI provides a stylistic benchmark: when asked an informal question about handling a talkative coworker, GPT-5.5 Instant used 30.2% fewer words and 29.2% fewer lines than GPT-5.3 Instant while delivering a response rated as more useful and appropriately toned.
This addresses a long-standing criticism from users who complained about ChatGPT's overly long, over-formatted, and excessively cautious responses.
Availability
Enhanced personalization from conversation history, files, and Gmail is rolling out to Plus and Pro users on the web and coming soon to mobile. Memory sources are rolling out across all ChatGPT consumer plans on the web. Expansion to Free, Go, Business, and Enterprise plans is planned for the coming weeks. Availability of specific personalization sources may vary by region.
Why this matters
GPT-5.5 Instant is the product used daily by more people than any other AI system. A reduction in hallucinations of over 50% on high-stakes prompts — if confirmed by independent testing — is a change with real practical significance wherever ChatGPT is used for medical, legal, or financial queries by people without specialist knowledge.
The memory sources mechanism is a step toward transparency in AI systems that increasingly operate on users' private data. Showing which data the model used and giving users the ability to delete it is not just a usability feature — it is a response to regulatory pressure around rights of data access and correction. The competitive context is relevant: Anthropic and Google are both intensively developing personalization in their own models and memory control systems. OpenAI, by deploying memory sources now, takes a leadership position in this layer, but the advantage is temporary.
What's next
- Enhanced personalization will reach Free, Go, Business, and Enterprise plans within the coming weeks — OpenAI has not provided a detailed timeline
- GPT-5.3 Instant will be retired three months after GPT-5.5 Instant's launch — developers using `chat-latest` in the API should verify their applications' compatibility
- OpenAI has announced further improvements to the memory sources view to provide a more complete picture of the factors influencing model responses
