GPT-5.5 Instant becomes default ChatGPT model with 52.5 per cent fewer hallucinations

GPT-5.5 Instant is now the default ChatGPT model for everyone, with OpenAI claiming a 52.5 per cent drop in hallucinated claims on high-stakes prompts and a jump from 65.4 to 81.2 on the AIME 2025 mathematics benchmark. OpenAI released GPT-5.5 Instant on 5 May, replacing GPT-5.3 Instant as the foundation model behind ChatGPT and as chat-latest in the API.

What GPT-5.5 Instant actually changes

GPT-5.5 Instant is the default model now visible to every ChatGPT user. The headline number is the 52.5 per cent reduction in hallucinated claims on what OpenAI calls high-stakes prompts, the kind of question you ask when the answer has to be right: dose advice in medicine, contract clauses in law, suitability rules in finance. That is the area where ChatGPT has done the most damage to its own credibility over the past two years, so a measurable drop here matters more than another general benchmark win.

The benchmark wins are still substantial. AIME 2025, a widely cited maths reasoning test, moves from 65.4 to 81.2, a jump of nearly 16 points. The multimodal MMMU-Pro score rises from 69.2 to 76, suggesting OpenAI has tightened image and chart understanding rather than just text reasoning. There is also a wider on-device strategy worth watching: the same week that GPT-5.5 Instant lands, Google’s own on-device push detailed in our coverage of Gemma 4 open weights shows how aggressively the cloud and edge sides of the AI conversation are now diverging.

GPT-5.5 Instant brings personalisation that actually behaves

The other meaningful change is personalisation. GPT-5.5 Instant can now reference past conversations, files you have uploaded and a connected Gmail account when answering. Memory sources are visible across all models, which means users can delete or correct stored facts rather than just toggling memory on and off. Shared chats do not expose those memory sources to the recipient, which closes one of the more obvious privacy holes in ChatGPT memory until now.

Plus and Pro subscribers on the web see the personalisation features first, with the mobile apps and the Free, Go, Business and Enterprise tiers following in the coming weeks. Developers calling chat-latest in the API are using GPT-5.5 Instant from launch day, and GPT-5.3 Instant remains accessible through model configuration for three months before retirement. This is the gentlest deprecation path OpenAI has shipped, perhaps remembering the backlash when GPT-4o was abruptly retired. For anyone tracking the broader assistant market, our ChatGPT vs Claude vs Gemini comparison sets the context.

GPT-5.5 Instant benchmark summary

Treat the OpenAI numbers as marketing-confirmed rather than independently verified, but the spread is convincing. The interesting test will be live-traffic hallucination rates measured by third parties over the next month, particularly on legal and medical prompts where the company has the most exposure. The release also lands in a week stacked with AI infrastructure news, including details from the OpenAI-Cerebras compute deal and the wider Anthropic-Amazon £79 (about $100)bn arrangement.

Conciseness and the GPT-5.5 Instant system card

The accuracy story comes with a tone story. OpenAI says GPT-5.5 Instant produces 30.2 per cent fewer words and 29.2 per cent fewer lines than GPT-5.3 Instant on like-for-like prompts, and the model is tuned to avoid “gratuitous emojis” and dramatic filler phrases. This is the second consecutive Instant update aimed at conversational tone: GPT-5.3 Instant landed on 3 March 2026 under OpenAI’s “more accurate, less cringe” banner, and GPT-5.5 Instant pushes the same dial further while adding a separate 37.3 per cent reduction in inaccurate claims on conversations users had previously flagged for factual errors. That second figure matters because it is measured on real failure cases rather than a curated benchmark set.

The published GPT-5.5 Instant system card is more candid than the headline release. OpenAI confirms statistically significant regressions versus GPT-5.3 Instant on two disallowed-content categories, gore and sexual content, while reporting comparable performance on violent, extremist, hate and self-harm prompts. It is also the first Instant-class model OpenAI has classified at “High capability” for biological and chemical content under its Preparedness Framework, which means automated monitors, age-gated protections and actor-level enforcement carry more of the safety load than the base model itself. UK users running ChatGPT in shared, family or workplace contexts should treat that as a reason to keep parental controls on and avoid pasting sensitive client information into shared chats, regardless of the headline accuracy gains.

What GPT-5.5 Instant means for UK ChatGPT users

For UK Plus subscribers paying about £20 a month, the practical change is straightforward: every default ChatGPT response from 5 May onwards comes from a noticeably more accurate model. The personalisation layer, however, is where UK buyers need to think harder. Connecting a Gmail account improves answers about your own data but also widens the surface area for content leakage if you share chats, run team workspaces or hand a phone to a child. The new memory-source visibility helps but only if you actually open the panel and review it.

Free and Go tier users in the UK should see GPT-5.5 Instant within the next few weeks, with Enterprise customers staged behind to allow admin policy controls to land first. Developers using the API for production workflows should test their evaluation suites against chat-latest before assuming feature parity with their pinned GPT-5.3 calls; OpenAI’s wording suggests instruction-following is broadly improved, which can also mean prompts written for the older model behave differently. For everyday consumer use, GPT-5.5 Instant is the most credible default ChatGPT model OpenAI has shipped, and the trust gap from late 2024 finally looks narrower.