ChatGPT’s New Default Model Cuts Hallucinations and Remembers Your Context

On 5 May 2026, OpenAI made GPT-5.5 Instant the new default model in ChatGPT, replacing the GPT-5.3 Instant that most people had been using. For the millions who reach for ChatGPT without ever choosing a model, this is the quiet kind of change that matters most: the everyday experience just shifted underneath them. Two things stand out: the model is designed to be wrong less often where it counts, and it can now draw on far more of your context to answer well.

What OpenAI actually changed

OpenAI says GPT-5.5 Instant reduces hallucination in sensitive areas such as law, medicine and finance, while keeping the low latency of its predecessor. In other words, the headline trade-off it is chasing is accuracy without making people wait longer for an answer. The company points to measurable gains on standard benchmarks as evidence: a score of 81.2 on the AIME 2025 maths test, up from 65.4, and 76 on the MMMU-Pro multimodal reasoning test, up from 69.2.

The second change is about context. GPT-5.5 Instant can use its search tool to refer back to past conversations, files and Gmail to give more personalised answers. ChatGPT now also shows its memory sources across all models, and you can delete or correct sources that are out of date. Notably, if you share a chat with someone, they will not be able to see those memory sources (a small but sensible boundary).

Who gets it, and when

The rollout is staged rather than instant. GPT-5.5 Instant arrives first for Plus and Pro users on the web, with free, Go Business and enterprise users following in the coming weeks, and mobile after that. For developers, GPT-5.5 is available through the API as chat-latest, while the previous 5.3 version stays accessible to paid users for three months. The practical takeaway is that what your staff see in ChatGPT this week may differ by plan and platform for a little while yet.

Why it matters for businesses

Two themes in this update map directly onto the questions businesses have been asking about AI. The first is trust. A model that is wrong less often in law, medicine and finance is precisely the kind of improvement that decides whether a tool is safe to use for real work rather than just brainstorming. It does not remove the need to check important answers (no model does), but it raises the floor on the domains where a confident mistake is most costly.

The second is context, and this is where the opportunity and the risk sit together. An assistant that can read your past chats, your files and your Gmail gives far more useful, tailored answers. But the moment a model reaches into your own data, the questions change from “is the answer good?” to “what can it see, who approved that, and where does the data go?” Visible, editable memory sources are a welcome step towards answering those questions in the consumer product, but inside a company, the same connection needs deliberate governance, not a personal toggle.

What to do about it

If your teams already use ChatGPT, treat this as a prompt to be intentional rather than to panic. Decide which work genuinely benefits from connecting business data, and which should stay general. Be clear about what staff may paste into a consumer tool and what they may not. And where context really does improve outcomes (support, operations, internal knowledge), consider connecting your own data through a controlled path rather than leaving each person to wire up their inbox ad hoc.

The direction of travel is plain. The default assistant is getting more accurate and more context-aware at the same time. The organisations that benefit will be the ones that pair that capability with clear rules about access and disclosure, so the gains in usefulness do not come at the cost of a leak.

Source: TechCrunch: OpenAI releases GPT-5.5 Instant, a new default model for ChatGPT