OpenAI on Monday, May 5, replaced the default model behind ChatGPT for every user. GPT-5.5 Instant supersedes GPT-5.3 Instant, and the launch headline is from OpenAI’s own evaluations: 52.5% fewer hallucinated claims on high-stakes prompts in medicine, law, and finance.
In English: ChatGPT, on the topics where confidently wrong actually hurts, is now confidently wrong about half as often. That’s OpenAI’s claim, anyway. It’s also the most consumer-relevant frontier-model improvement in the past twelve months, because the things consumers ask ChatGPT — what should I do about this rash, can I be fired for this, how do I pay this tax — are exactly the categories the eval covers.
The update is part of a four-pack OpenAI rolled out the same day. The other three:
- A realtime voice intelligence model in the API that can reason, translate, and transcribe inside a single inference path. Translation: the gap between “voice assistant” and “actually useful conversation partner” just got smaller.
- A personal finance preview inside ChatGPT for U.S. Pro subscribers. Connect your accounts, see a dashboard, ask questions grounded in your real money. Privacy implications: you decide.
- Codex Mobile in preview, with extended SSH, hooks, and HIPAA support for enterprise. Developers can now babysit their coding agents from a phone.
OpenAI also dropped a bombshell that nobody is sure how to evaluate yet: the lab claims one of its models disproved a central conjecture in discrete geometry. The mathematics community has not weighed in. The launch communications were thin on detail.
For everyday ChatGPT users, the practical implication is: the answers should be better, starting now. You don’t have to do anything; the default has changed under you. You may notice the chatbot is meaningfully less likely to confidently invent things in the categories where it used to. Or you may not notice, because, well, most people only notice the model failing.
Two weeks later, at Google I/O on May 19, Google countered with Gemini 3.5 Flash, the new Omni multimodal model, and a general-purpose agent named Spark. The frontier labs are now firing on a roughly two-week cadence. The category is no longer “AI assistant in your browser tab.” The category is now “ambient AI layer in your operating system.” OpenAI is competing for that consumer attention without owning the OS. Google is competing for it while owning the OS.
The next two quarters tell you who’s winning. ChatGPT just got better by default. That’s the move.