OpenAI ships GPT-5.5 Instant

OpenAI has released GPT-5.5 Instant as the new default model in ChatGPT, replacing GPT-5.3 Instant and tightening the company’s already rapid product cycle. TechCrunch reports that OpenAI is pitching the update as faster while reducing hallucinations in high-stakes domains such as law, medicine and finance.

The change is small in name but large in operational meaning: “default” determines what most users experience, what developers benchmark against, and what competitors must match to avoid looking slow or unreliable. OpenAI is also using the release to normalise frequent forced migrations. GPT-5.3 will remain available via API for paid users for three months, TechCrunch says, after which developers either absorb the switch or start paying the integration cost of maintaining legacy behaviour elsewhere.

OpenAI’s own numbers, as relayed by TechCrunch, are designed to make the upgrade feel non-optional. GPT-5.5 Instant scored 81.2 on the AIME 2025 math test versus 65.4 for GPT-5.3, and 76 versus 69.2 on MMMU-Pro for multimodal reasoning. Whether or not those benchmarks map neatly onto day-to-day chat, they provide a simple narrative for procurement teams: buy the newer model or explain why you chose the one that scores lower.

The more strategically important shift is how OpenAI is expanding “memory” and retrieval inside ChatGPT. GPT-5.5 Instant can use a search tool to refer back to past conversations, files and Gmail to deliver personalised answers, with rollout starting on the web for Plus and Pro users and expanding to other tiers later, according to TechCrunch. OpenAI also plans to show “memory sources” across models—an attempt to make personalisation auditable enough that users can delete or correct what the system is drawing from.

That feature set changes the competitive moat from raw model quality to account-level lock-in. A chatbot that remembers your documents and email becomes less interchangeable, even if rival models are similar on public benchmarks. It also creates a new class of failure: if an answer is wrong because an old file was indexed or a prior chat was misunderstood, the error is now tied to a personal data trail rather than a generic “hallucination.”

OpenAI has already learned that model churn has a user-experience cost. TechCrunch notes the backlash when OpenAI withdrew GPT-4o in February 2026, after some users said they felt emotionally attached to its “personality.” GPT-5.5 Instant is framed as a performance upgrade, but it still moves millions of people onto a different system response surface overnight.

ChatGPT’s default model has changed again, and the previous default now comes with a three-month countdown for API users.