52.5 percent. That's the drop in hallucinated claims OpenAI reports after replacing ChatGPT's default model with GPT-5.5 Instant. The company says the update also cuts inaccurate claims by 37.3 percent on conversations previously flagged by users and improves visual reasoning, STEM performance and web-search judgement. The change arrives with new memory-source controls and desktop app updates designed to make the assistant faster and more transparent for the hundreds of millions who use it.

Hundreds of millions of ChatGPT users will get fewer incorrect answers because OpenAI has made GPT-5.5 Instant the default model across its service. The company frames the rollout as a quality and efficiency upgrade for the "daily driver" model that most people encounter, and it sets a new baseline for how the assistant handles factual tasks and contextual memory.

Accuracy gains and where they matter

OpenAI's release notes say GPT-5.5 Instant reduces hallucinated claims by 52.5 percent versus GPT-5.3 Instant on a set of high-stakes prompts covering medicine, law and finance. The notes add the model cuts inaccurate claims by 37.3 percent on conversations that users had previously flagged for factual errors. Those are specific, measurable improvements rather than general claims of being "better"; OpenAI also reports internal gains on visual reasoning, mathematics and science benchmarks, and says the model makes more reliable decisions about when to run a web search to strengthen an answer.

Shorter, less verbose replies are a deliberate goal. OpenAI says GPT-5.5 Instant aims to give briefer answers without losing substance and to ask fewer unnecessary follow-up questions. That, combined with improved context retrieval from past chats and attached files, is meant to reduce the need for users to repeat themselves across sessions.

Memory sources and user control

One of the most visible changes is a feature OpenAI calls memory sources. The interface shows which contextual items the model used when generating a response, for example saved memories, prior conversations, files attached to a chat or content from a connected Gmail account. The company says those citations are visible only to the chat owner when a response is shared, and provides controls so users can delete or correct any item the model cited.

OpenAI reiterates its existing opt-out for using user content to train public models and confirms that enterprise inputs supplied via paid offerings aren't used to improve model performance. The company emphasises that users remain in control of stored context, presenting the memory sources feature as both a transparency tool and a corrective mechanism for personalised replies.

Alongside the memory sources UI, OpenAI notes the model now uses context from past chats, files and connected Gmail more effectively. The firm describes these changes as aimed at both improving relevance and lowering repeat queries rather than adding new capabilities alone.

Mastheads

Regional coverage of product integrations flagged several convenience features arriving alongside the model update. Those reports note a browser extension that lets people query ChatGPT from the address bar and an image-generation capability that gives non-subscribers a limited number of free creations per day. Canvas and Projects continue to be features targeted at paying subscribers for collaborative work and organisation.

Desktop apps and privacy details

Client-side updates come with the model change. TechSpot summarised platform notes showing new launchers on desktop for quick queries and screenshot-based interactions, with the Windows launcher keyed to Alt + Space and the macOS launcher keyed to Option + Space. The macOS app includes native voice input and deeper native integration, while the Windows build more closely wraps the web experience.

TechSpot's notes identify the Windows desktop package by version number 2026.430.708.0.

TechSpot's app documentation also states that files uploaded to a conversation are retained while the conversation is active and for three hours after a conversation is paused, after which those files are deleted. Deleted chats are removed within 30 days unless they're de-identified or retained for legal or security reasons. Those retention details matter for users who share sensitive documents during a session and want predictable deletion timelines.

OpenAI's documentation and platform notes provide step-by-step instructions for privacy controls, training opt-outs and the new memory sources interface. The company positions these tools as part of a shift from opaque personalisation toward greater visibility and user agency over how context is stored and used.

Functionally, the update touches how users start and conduct conversations. On desktop and mobile, people can upload photos to begin chats and discuss screenshots directly inside the app. Keyboard shortcuts and voice input are intended to speed simple queries and screen-based work. Better context retrieval should also reduce the temptation to paste the same background information into multiple chats.

OpenAI describes GPT-5.5 Instant as a performance-focused update rather than an experimental leap. The model's improvements are framed as an operational refinement to the core experience the company expects most users to rely on every day.

That framing is important. By making the improved model the default, OpenAI narrows the gap between the versions used by casual users and those available to power users, while keeping enterprise data protections and training opt-outs intact. For organisations and individuals who care about accuracy on technical topics, the stated percentage reductions give a concrete metric to track.

Related Articles

The most tangible detail from the rollout is the Windows desktop package, identified as version 2026.430.708.0, which carries the new features.

This article was created with AI assistance.