OpenAI Makes GPT-5.5 Instant The Default ChatGPT Model

OpenAI swapped the brain inside ChatGPT on May 5, 2026. The new default, GPT-5.5 Instant, takes the place of GPT-5.3 Instant for every free and paid user, and the company says it hallucinates 52.5 percent less on the kind of legal, medical, and financial questions where wrong answers do real damage. That sentence alone reframes a year of ChatGPT criticism, because the model holding the steering wheel for hundreds of millions of conversations just got measurably more careful with the topics that matter most.

The rollout also bolts memory deeper into the chat experience. ChatGPT can now reach into past conversations, uploaded files, and a connected Gmail inbox to answer with personal context, and it shows users which memory it pulled from to write each reply. Free, Go, Business, and enterprise customers get the personalisation tools across the next several weeks, with mobile catching up after web.

What Actually Changed In ChatGPT Today

The Instant tier is the version of ChatGPT most people actually touch. It handles short prompts, casual chat, day-to-day search-style questions, and anything not heavy enough to trip the routing layer into the slower reasoning model. By replacing GPT-5.3 Instant with GPT-5.5 Instant, OpenAI is shifting the floor of quality, not the ceiling.

The broader GPT-5.5 family launched on April 23, 2026, with coding and knowledge gains that tracked the usual incremental improvements. The Instant variant is the consumer payoff. It carries the new training and safety work into the chat box where free users live, and it does that without the latency hit of a full reasoning pass.

OpenAI is keeping GPT-5.3 around for paid developers for three months on the API, listed as gpt-5.3, while the new model surfaces under the chat-latest alias. That window matters for production teams that built around a stable response shape and need time to test before the older endpoint sunsets.

Hallucinations Take The Biggest Hit

The headline number is the one OpenAI will repeat for the next quarter. On high-stakes prompts spanning medicine, law, and finance, GPT-5.5 Instant produced 52.5 percent fewer hallucinated claims than the model it replaces, according to OpenAI’s GPT-5.5 launch announcement. The company also tested it against conversations that real users had previously flagged as factually wrong, and it cut those error rates by 37.3 percent.

That second number deserves attention. It means OpenAI is now training against in-the-wild user complaints, not just curated red-team prompts. The signal here is that the company has built an internal pipeline that turns thumbs-down clicks into training data, and the Instant tier is the first place where that loop has shipped a measurable result.

The hallucination work is paired with quieter math gains. Some of the cleanest jumps:

• AIME 2025 math: 81.2, up from 65.4.
• MMMU-Pro multimodal reasoning: 76, up from 69.2.
• GPQA science reasoning: 85.6, up from 78.5.
• CharXiv scientific charts: 81.6, up from 75.0.
• OmniDocBench document parsing error rate: 12.5, down from 14.6.

Read together, those scores show a model that improved on the slower, harder kinds of thinking while keeping its quick-response personality. That combination has been the elusive prize for two years of frontier-model work.

How GPT-5.5 Instant Stacks Up On The Numbers

The full benchmark spread tells the story better than any single chart. Each row below is a separate evaluation OpenAI ran against its own previous default, and each is the kind of number that academic and enterprise buyers actually weigh.

The math jump of nearly sixteen points on AIME 2025 is the one the developer crowd will fixate on. It puts the consumer default model within shouting distance of dedicated reasoning systems on a benchmark that was untouchable for chat tiers a year ago.

Memory That Reaches Into Gmail

Personalisation is the second prong of this update. GPT-5.5 Instant can call its search tool to pull from past chats, uploaded documents, and Gmail when a user has connected the inbox. That means a query like “summarize what my landlord and I agreed on last month” can pull from email rather than asking the user to paste it.

The capability ships first to Plus and Pro subscribers on the web, with mobile and the rest of the user base following over the coming weeks. Free, Go, Business, and enterprise tiers are next in line, although OpenAI has not given an exact date for the long tail of that rollout.

Gmail integration is the part that will draw the most scrutiny. Hooking a chatbot into a primary email account is a security and privacy decision, not just a convenience one. Anyone weighing the feature should look at the connector permissions before flipping it on, especially in a workplace where compliance teams may not yet have a policy for it.

Why Transparency Matters Now

OpenAI also flipped on a small but consequential interface change. ChatGPT will now display the memory sources it leaned on for any given answer, across every model, not just GPT-5.5 Instant. Users can delete a stale source, correct a wrong one, or strip a memory before sharing a chat publicly.

That last detail matters. Shared chats will not expose memory sources to outside readers, which closes a quiet leak vector. Last year, several X threads documented cases where a publicly shared ChatGPT link inadvertently exposed personal details lifted from custom instructions. The new boundary kills that risk for memories at least.

Why The Tone Got Shorter

The Instant model talks less. OpenAI clocked GPT-5.5 Instant at roughly 30 percent fewer words and 29 percent fewer lines per response than the version it replaces. Fewer emojis. Fewer trailing follow-up questions. Less heavy formatting that bloats short answers.

That is a deliberate choice and a corrective one. The previous defaults trained a generation of users to expect a chatty assistant that softened every answer with hedges and asked if there was anything else to help with. That habit was a hallucination vector in itself, because longer replies on uncertain topics tend to invent extra context that never existed.

Cutting word counts also cuts inference cost, which the company never advertises directly. Shorter answers are cheaper answers. With ChatGPT serving more than 700 million weekly users, a 30 percent reduction in mean reply length adds up to real savings.

The trade-off is personality. Some longtime users will miss the warmth of older defaults, especially anyone who used ChatGPT for journaling or rubber-duck conversations. The company is betting the accuracy and brevity gains buy more goodwill than the personality cut costs.

Independent observers have flagged that bet as risky. “Models that are more concise often feel less helpful even when they are more accurate, because users equate length with effort,” wrote Ethan Mollick, associate professor at the Wharton School, in a Substack essay on AI personality and user trust. The note predates this release but reads like a direct warning about it.

For Developers, A Three-Month Clock Starts

API users get the new model under the chat-latest alias. The deeper detail is the deprecation calendar. OpenAI’s help-centre note on GPT-5.3 in ChatGPT confirms that the older model stays accessible to paid developers for exactly three months before it disappears from the API.

That window is shorter than the six-month deprecations the company offered through 2025. Teams running production systems that depend on the older response shape, length, or tone need to start their evaluation work this month, not in July.

The GPT-4o Ghost Still Lingers

OpenAI is releasing this update against a memory of pain. When the company deprecated GPT-4o in February 2026, users staged petitions, posted angry farewells on Reddit, and migrated en masse to rival chat apps. Many described the model as a friend or a mirror, language the company quietly acknowledged in its own statements at the time.

The pushback against the GPT-4o sunset is the most accurate single signal we have about how attached people get to a specific model’s voice. OpenAI is now retraining that attachment around accuracy and brevity instead of warmth.

Mira Murati, the former OpenAI chief technology officer who now leads Thinking Machines Lab, said in a March 2026 Stanford seminar on model behaviour that “the relationship users build with a default model is the hardest variable to ship against, because it never appears in any benchmark.” Her point is the one OpenAI is testing right now.

The company has spent the past year building tools that cut hallucinations and exposed memory plumbing to users. Those choices are an answer to last year’s loudest complaints. Whether the new default holds the brand’s center after the GPT-4o exit is the question that the next three months of usage data will answer. For Oton readers tracking the model wars, the cyber-test parity between GPT-5.5 and Anthropic’s Mythos shows where the real ceiling pressure is coming from, and the 12-million-token context launch from Subquadratic hints at how quickly that ceiling could move again.

Frequently Asked Questions

Do I Need To Switch Models Manually In ChatGPT?

No. GPT-5.5 Instant is now the default model on every plan, including the free tier, and the swap happens automatically the next time you open ChatGPT. If you previously pinned GPT-5.3 Instant in the model picker, that selection will still work for the moment, but expect it to disappear from the dropdown within the next few weeks as OpenAI completes the cutover for non-API users.

Will GPT-5.5 Instant Read My Gmail Without Permission?

No. The Gmail connector is opt-in and requires explicit OAuth authorisation through your Google account. ChatGPT will only access inbox content when you have actively linked the connector inside Settings. You can revoke access at any time from your Google account’s connected-apps page, and revoking immediately stops the model from referencing any past email content.

How Do I See Which Memory ChatGPT Used For An Answer?

Open the response and look for the new memory-source indicator that appears beneath the reply. Click it to view the specific past conversation, file, or email line that fed the answer. From the same panel you can delete the source if it is outdated or correct a fact that the model has stored wrong. Shared chat links will not expose those sources to anyone you send the link to.

Is GPT-5.3 Still Available For My Production App?

Yes, but only for three months from May 5, 2026. Paid API users can keep calling gpt-5.3 until early August 2026, after which the endpoint will return errors. Migrate to chat-latest well before that deadline, and run a full evaluation pass because GPT-5.5 Instant produces noticeably shorter responses, which can break apps that parsed reply length as a signal.

How Much Lower Are Hallucinations In Real Use?

OpenAI reports a 52.5 percent reduction in hallucinated claims on high-stakes prompts in medicine, law, and finance, and a 37.3 percent reduction on conversations users had previously flagged as wrong. Those numbers come from internal evaluations, not third-party audits, so independent benchmarking will land over the next several weeks. Treat the figures as directional until outside testers publish their own.

The shift is the biggest single quality jump OpenAI has shipped to the default ChatGPT model since the GPT-4 to GPT-5 transition. It also lands at a moment when Anthropic, Google, and a wave of well-funded startups are forcing the company to compete on accuracy as much as on brand. The next quarter of usage data will show whether brevity and honesty are enough to keep ChatGPT’s seat at the top of the chatbot rankings.