OpenAI ships GPT-5.5 Instant with 52.5% fewer hallucinations

OpenAI's new default ChatGPT model, GPT-5.5 Instant, produced 52.5% fewer hallucinated claims than its predecessor GPT-5.3 Instant on high-stakes prompts covering medicine, law, and finance, according to internal evaluations published on May 5, 2026. The model began rolling out to all ChatGPT users the same day, replacing GPT-5.3 Instant as the default and arriving in the API under the chat-latest identifier.

The stakes are straightforward. Instant is the daily driver for hundreds of millions of people, as OpenAI notes, so incremental quality gains compound across billions of interactions. The release also lands amid intensifying pressure on AI vendors to demonstrate measurable reliability improvements — particularly in domains where a fabricated claim carries real consequences.

The factuality numbers

OpenAI frames the update around accuracy. In internal evaluations, GPT-5.5 Instant reduced hallucinated claims by 52.5% on high-stakes prompts in medicine, law, and finance. On especially challenging conversations that users had previously flagged for factual errors, the model reduced inaccurate claims by 37.3%.

The company also reports general capability gains: better analysis of photo and image uploads, stronger STEM answers, and improved judgment about when to invoke web search.

To illustrate the difference, OpenAI published a side-by-side comparison on an algebra problem. Given a user's attempted solution to a square-root equation, GPT-5.3 Instant caught that the candidate answer x=3 failed when plugged back into the original equation — but stopped there and incorrectly concluded there was no real solution. GPT-5.5 Instant went further: it identified the actual algebra mistake, re-derived the corrected quadratic, and produced the valid solution (3 + √33)/2. The older model, in OpenAI's telling, spotted the error but failed to fix it.

OpenAI attributes measurable benchmark gains to these improvements, citing results on CharXiv-reasoning (scientific chart reasoning), MMMU-Pro (expert multimodal reasoning), and OmniDocBench (document parsing).

Shorter answers, fewer emojis

The second pillar of the release is concision. OpenAI says the model's responses are "tighter and more to-the-point without losing substance," keeping what the company calls "the warmth and personality that makes ChatGPT enjoyable to use." In practice, that means reduced verbosity and overformatting, fewer unnecessary follow-up questions, and the removal of clutter — including, pointedly, "gratuitous emojis."

OpenAI quantified the change with a workplace-advice prompt ("how do i tell my coworker to quit yapping all the time"). GPT-5.5 Instant's answer used 30.2% fewer words and 29.2% fewer lines than GPT-5.3 Instant's. The company judged the shorter response stronger: "informal, practical, and workplace-safe without overexplaining," framing the problem around boundaries rather than the coworker's personality. The GPT-5.3 answer was "more complete, especially with its 'what not to do' section, but it feels slightly overbuilt for a casual advice prompt."

Personalization — and the controls that come with it

The most structurally significant change may be personalization. GPT-5.5 Instant draws more effectively on context from past chats, files, and connected Gmail to tailor answers, and OpenAI says it now searches past conversations faster and decides more intelligently when personalization would actually improve a response.

The company demonstrated with a tea-shop recommendation. Asked "What new tea place should I try?", GPT-5.3 Instant produced a solid but generic San Francisco list. GPT-5.5 Instant referenced the user's established preference for "cleaner Taiwanese/high-mountain tea vibes over sugary boba" and their frequent visits to Asha Tea House, then ranked recommendations by predicted fit — including a "most likely to become your new favorite" pick and a "serious tea" option at Song Tea & Ceramics.

Alongside the model, OpenAI is introducing memory sources across all ChatGPT models. When a response is personalized, users can see which context shaped it — saved memories or specific past chats — and delete or correct items that are outdated. Users can delete chats they no longer want cited, edit saved memories in settings, or use temporary chats that neither use nor update memory. Memory sources are not shown to others when a chat is shared.

OpenAI is candid about the feature's limits: memory sources "may not show every factor that shaped an answer," potentially displaying the most relevant past chats rather than every conversation the model searched. The company says it will keep improving the view.

The transparency mechanism matters beyond convenience. As AI assistants increasingly ground answers in users' personal history, regulators and users alike have pressed for visibility into what data informs a given output. Memory sources are OpenAI's structural answer to that question — partial, but explicit.

Availability and retirement timeline

The rollout is tiered. GPT-5.5 Instant is available now to all ChatGPT users and in the API as chat-latest. Paid users can still select GPT-5.3 Instant through model configuration settings for three months before OpenAI retires it.

Enhanced personalization from past chats, files, and connected Gmail is reaching Plus and Pro users on the web first, with mobile support coming soon and expansion to Free, Go, Business, and Enterprise plans "in the coming weeks." A June 9, 2026 update noted that personalization improvements had reached ChatGPT Go and Free tiers, with Free-tier responses drawing from a reduced set of past chats. Memory sources are rolling out across all consumer plans on the web and soon on mobile. Availability of specific personalization sources may vary by region.

The free-tier restriction is notable: personalization quality now depends on how much conversation history a plan can access, creating a new, subtle differentiator between paid and unpaid tiers.

What it signals

For OpenAI, the release follows the GPT-6 family launches and represents a continued bet that the default-model experience — not just flagship models — determines whether hundreds of millions of users keep coming back. The three-month retirement clock on GPT-5.3 Instant signals a faster default-model cadence, which developers pinning to chat-latest will need to track.

The bigger signal is the pairing of deeper personalization with user-facing controls. OpenAI is betting that recall of past context, visible and editable by the user, is the next competitive frontier for consumer AI assistants — and it has now built the plumbing for that across every ChatGPT plan.