OpenAI has replaced GPT-5.3 Instant with GPT-5.5 Instant as the default model for all ChatGPT users, free and paid alike, with no price change attached. The headline claim is a 52.5% reduction in hallucinated statements across medicine, law, and finance — the exact domains that UK regulators like the FCA and MHRA are currently stress-testing for AI accuracy. For anyone who has ever caught ChatGPT confidently making something up, this is the update worth paying attention to.
The accuracy claim
OpenAI says the new model also produces 37.3% fewer inaccurate responses during long, complex conversations — the kind where previous models would quietly lose the thread. Responses are shorter too: 30.2% fewer words and 29.2% fewer lines, with the gratuitous emoji and bullet-point padding stripped out.
On benchmarks, GPT-5.5 Instant scores 81.2% on AIME 2025 (a competitive math test), up from 65.4% for its predecessor, and 85.6% on GPQA, a PhD-level science evaluation. Those are meaningful jumps for what is still the entry-level, always-available model.
One caveat worth stating plainly: all these figures come from OpenAI’s own internal evaluations. Independent third-party benchmarks have not yet confirmed them. External testing is expected within the next few weeks, per TechCrunch.
Memory sources
The other notable addition is a feature called Memory Sources. ChatGPT can already remember details across conversations; now it tells you which past chats, uploaded files, or connected Gmail data shaped a given response. You can inspect that list, delete specific entries, or correct them. When you share a chat with someone else, the source list stays hidden — only you see it.
This transparency layer has immediate relevance for UK users. The Information Commissioner’s Office has been pressing AI companies on exactly this kind of personal-data disclosure. The new UI gives users a testable way to check what the model knows about them — and to remove it.
UK Business and Enterprise users should note that Gmail and file-based personalization is rolling out web-first for Plus and Pro subscribers immediately, with Free, Go, Business, and Enterprise tiers following over the coming weeks. Some regional variation in that timeline is expected.
What this means day-to-day
For most people, the upgrade is invisible — ChatGPT just gets a bit more reliable without any action needed. The conciseness improvement alone removes one of the most common complaints: responses padded out with unnecessary context and formatting. The hallucination reduction, if it holds under independent scrutiny, would make the free tier meaningfully more trustworthy for quick research in sensitive areas — without requiring an upgrade to a paid plan.