OpenAI drops GPT-5.5 Instant as ChatGPT's new default

If ChatGPT is your daily driver, you might notice things feel a bit cleaner this week. The filler is gone, the random emojis are missing, and most importantly, it seems to have stopped making things up quite as often.

OpenAI just swapped out ChatGPT's default engine for GPT-5.5 Instant, rolling it out to all users as of May 5. Following last month's GPT-5.5 launch, this brings the latest architecture directly into the lightweight model hundreds of millions talk to every single day.

Halving the hallucinations

The biggest upgrade comes where AI is typically the most dangerous: facts. In internal tests on high-stakes topics like medicine, law, and finance, OpenAI says GPT-5.5 Instant produced 52.5% fewer hallucinated claims than GPT-5.3 Instant. On tricky conversations that users had previously flagged for factual errors, inaccurate statements dropped by 37.3%.

It is also flexing much better reasoning under the hood. The model scored 81.2 on the AIME 2025 math exam, up from 65.4 on the older version, and jumped to 76 from 69.2 on the MMMU-Pro multimodal benchmark.

Beyond raw test scores, day-to-day conversation is noticeably tighter. OpenAI dialed back the unnecessary formatting, dropped the endless follow-up prompts, and cut down on gratuitous emojis. It also makes sharper calls on when to browse the web instead of bluffing its way through.

Connecting your inbox with receipts

Personalization is getting a massive overhaul, too. The model can now search through your past chats, uploaded documents, and connected Gmail to deliver answers tuned to your life. The feature is starting with Plus and Pro users on the web, with free, business, and mobile tiers following soon.

Because digging through personal emails naturally raises eyebrows, OpenAI is pairing this with a new feature called memory sources across all models. When ChatGPT personalizes an answer, it now shows you the exact chat or saved note it pulled from. You can delete or correct anything outdated, and none of those source logs are exposed if you share a chat link.

For developers, the model arrives in the API as chat-latest. Paid ChatGPT subscribers can still fall back to GPT-5.3 Instant for the next three months. OpenAI is clearly being careful after the massive user outcry over retiring GPT-4o, which fans petitioned to save as their "best friend" before it was officially shut down in February 2026.

Why it matters

This update is a masterclass in practical engineering. Instead of chasing vanity benchmark jumps, OpenAI targeted the three things that actively annoy daily users: bad answers in critical moments, excessive verbosity, and the creepy feeling of an unpredictable personalization engine.

For regular users, cutting hallucination rates in half on high-stakes topics makes ChatGPT significantly more trustworthy as a work assistant. The added transparency of memory sources gives users the steering wheel: if the AI remembers an old project incorrectly, you can simply wipe that memory clean rather than battling it in future prompts.

Strategically, tying Gmail and chat histories directly into real-time answers puts OpenAI in direct combat with Google Gemini's workspace integrations. We think the battleground has firmly shifted from who has the biggest raw model to who can build the most dependable, low-friction operating system for your digital life.