OpenAI Launches GPT-5.5 Instant as New ChatGPT Default Model

OpenAI shipped GPT‑5.5 Instant on Tuesday, swapping out GPT‑5.3 as the default model powering ChatGPT for hundreds of millions of users. The new foundation model targets a specific pain point: hallucinations in high‑stakes domains like law, medicine, and finance, while keeping the snappy response times users expect from the “Instant” line, according to TechCrunch.

This isn’t a niche research model — it’s the engine that most people interact with every day. OpenAI released the full GPT‑5.5 family ¹ with promises of better coding and knowledge work, but the Instant variant is what actually ships to the masses.

The numbers tell a clear story. GPT‑5.5 Instant scored 81.2 on AIME 2025, a challenging math benchmark — up from 65.4 for the outgoing GPT‑5.3 Instant. On MMMU‑Pro, which tests multimodal reasoning across images and text, it hit 76 versus 69.2 for its predecessor, per.¹

One feature that stands out for builders: GPT‑5.5 Instant can now search across past conversations, uploaded files, and Gmail to ground its answers in your personal context. This is live for Plus and Pro subscribers on the web, with mobile coming soon. Free, Go Business, and enterprise accounts get it in the “coming weeks,” as reported by TechCrunch.

OpenAI also added memory source transparency: ChatGPT now shows which specific memories informed each answer, across all models. Users can delete bad memories or correct wrong ones. And if you share a chat link, the recipient can’t see your memory sources — a privacy detail that matters for anyone building products that integrate with ChatGPT’s conversation history.

For developers hitting the API, the GPT‑5.5 model is available as “chat‑latest” — the same dynamic alias OpenAI uses to roll new models into existing integrations. If you’re still building against 5.3, you have three months before it’s deprecated for paid users, according to TechCrunch.

The three‑month window is important context. When OpenAI yanked GPT‑4o earlier this year, the backlash was intense — users signed petitions, called the model their “best friend,” and described it as “a mirror” of their personality. OpenAI deprecated it anyway in February 2026.

GPT‑5.5 Instant follows a pattern: ship the big model (GPT‑5.5 family in April), then a month later ship the optimized Instant variant that becomes the default. It’s a two‑step launch cadence — power users and developers get early access to the full model via API, then the masses get a distilled version tuned for speed and cost.

For builders, the key signal is the memory architecture. Giving models access to Gmail, files, and conversation history isn’t just a feature — it’s OpenAI building the scaffolding for an AI that’s persistent across sessions, not just smarter within a single chat. That’s the kind of infrastructure that makes AI agents genuinely useful, not just impressive in demos.

Every new model launch is also a sunset announcement for something else. GPT‑5.3 Instant now joins GPT‑4o in the graveyard. The three‑month grace period is generous by OpenAI standards, but the pattern is clear: if you build on a specific model version, you’re on a clock.

For teams building production systems, this means API calls to named models are inherently temporary. The “chat‑latest” alias abstracts the churn but sacrifices control. Builders who need deterministic behavior should pin to specific dated snapshots — and budget for migration work every quarter.