OpenAI swaps in GPT-5.5 Instant as ChatGPT’s default, promising fewer made-up answers and clearer memory controls

OpenAI is changing the engine that powers ChatGPT for most users, and it’s betting the upgrade will make the chatbot less likely to confidently spit out wrong information.

Starting May 5, 2026, the company began rolling out GPT-5.5 Instant as ChatGPT’s new default model, replacing GPT-5.3 Instant. OpenAI says the new model delivers shorter, more reliable responses and introduces a new transparency feature that shows when ChatGPT is drawing on saved “memory” to personalize what it tells you.

The rollout is gradual, meaning two people with the same plan might not see the same model on the same day. Developers will also see the change through OpenAI’s API under the aliaschat-latest, a moving target designed to always point to the newest default chat model.

A new default model, without users touching a setting

The biggest practical shift is simple: GPT-5.5 Instant is now the model most people will use automatically, whether they realize it or not. That matters because default settings shape how millions of users experience ChatGPT, students, office workers, programmers, and anyone leaning on the tool for quick answers.

OpenAI is keeping GPT-5.3 Instant available to paying subscribers for about three months before retiring it. That grace period could matter for people and companies whose workflows depend on the older model’s style, especially if they’ve built prompts, templates, or internal tools that expect a certain format.

The update also underscores how fast OpenAI is iterating on its most-used product layer. GPT-5.5 Instant arrives roughly two months after GPT-5.3 Instant, skipping a “5.4” Instant release entirely.

OpenAI claims a 52.5% drop in hallucinations for high-stakes topics

OpenAI’s headline claim: GPT-5.5 Instant produces52.5%fewer “hallucinated” assertions than GPT-5.3 Instant on high-stakes prompts inmedicine, law, and finance. Those are internal tests, and OpenAI hasn’t published full methodology or the remaining error rate, so the numbers are hard to independently verify.

The company also says inaccuracies fell37.3%in conversations that users flagged for factual errors. That metric is closer to real-world use, where people report bad answers across messy, everyday scenarios, like tax questions, medication dosing, or contract language.

OpenAI is also highlighting improved self-correction. In one example, the company says GPT-5.5 Instant can spot a trap in a math problem, identify the faulty step, and redo the calculation instead of bulldozing ahead with a polished-but-wrong solution.

Still, “fewer hallucinations” doesn’t mean “no hallucinations.” Cutting errors in half can be a big deal, and still leave plenty of room for costly mistakes if users treat the output as authoritative in sensitive situations.

“Memory Sources” aims to show what ChatGPT remembered, and let you delete it

The most visible new feature is calledMemory Sources. When ChatGPT personalizes an answer using saved memory, OpenAI says users will be able to open a panel showing which stored details influenced the response, like a saved note or information pulled from a prior conversation.

Crucially, users can edit, mark as irrelevant, or delete those memory items. It’s a direct response to a common complaint about AI assistants: if the tool “remembers” you, people want to knowwhatit remembers and how to control it.

OpenAI is also adding atemporary chatmode that doesn’t read from memory and doesn’t write new memory. The pitch is straightforward: ask sensitive questions, run tests, or draft delicate messages without feeding your long-term profile.

There’s a catch. OpenAI says Memory Sources won’t necessarily show every factor behind an answer, only certain retrieved conversations may appear, with the view expected to become more complete over time. In other words: more transparency, not total traceability.

Shorter answers, fewer emojis, and a different feel

OpenAI says GPT-5.5 Instant is designed to be more concise: less filler, fewer unnecessary follow-up questions, lighter formatting, and fewer emojis. That may sound minor, but for heavy users, shaving even a couple paragraphs off routine responses can mean less scrolling and faster decision-making.

Concise output also plays better with workplace reuse. If you ask for “three options and pros/cons,” you want a clean list you can paste into a ticket, memo, or slide, not a long preamble.

OpenAI also says personalization will be “better judged,” meaning the model should be more selective about when memory actually helps. The risk, though, is that shorter answers can omit important caveats, especially in legal or medical contexts where exceptions and contraindications matter.

Benchmark scores are up, but real-world reliability is the real test

OpenAI and tech press coverage have pointed to benchmark gains as supporting evidence. TechCrunch cited a score of81.2onAIME 2025(a math-focused benchmark) versus65.4for GPT-5.3 Instant.

OnMMMU-Pro, a multimodal reasoning benchmark, GPT-5.5 Instant reportedly scored76versus69.2for its predecessor, suggesting improvements when the model has to reason across different kinds of inputs, not just text.

But benchmarks don’t guarantee better performance on your specific tasks, data, or edge cases. And because “Instant” models are tuned for speed and low latency, there’s still a tradeoff between quick responses and deeper, more deliberate reasoning, something users may notice depending on what they ask ChatGPT to do.

The bigger implication is that OpenAI is trying to make the default ChatGPT experience feel less like a clever chatbot and more like a dependable daily tool. Whether GPT-5.5 Instant earns that trust will come down to the moments that matter most: when the model is tempted to guess, and whether it knows when to stop.