OpenAI Swaps ChatGPT Default to GPT-5.5 Instant, Cuts Hallucinations 52.5%

OpenAI on Tuesday replaced ChatGPT's default model with GPT-5.5 Instant, a successor to GPT-5.3 Instant that the company claims produces 52.5% fewer hallucinated claims on high-stakes prompts in medicine, law and finance. The release, announced in a company blog post, also introduces a "memory sources" panel that surfaces some of the past chats, files and Gmail items ChatGPT pulled into a personalized response. The company is rolling the model out to all ChatGPT tiers and to the API, with paid users retaining access to GPT-5.3 Instant for three months.

Accuracy claims rest on internal evaluations

The 52.5% hallucination reduction comes from OpenAI's own evaluations on prompts the company classified as high-stakes. On a separate set of conversations that users had previously flagged for factual errors, inaccurate claims fell 37.3%, OpenAI said. OpenAI also disclosed a HealthBench score of 51.4 out of 100, up from 49.6 for GPT-5.3 Instant, and 38.4 on HealthBench Professional, the clinical version of the benchmark, up from 32.9. SiliconANGLE noted that the comparisons are limited to OpenAI's predecessor model and do not include rival systems from Google or Anthropic.

Per TechCrunch's writeup, GPT-5.5 Instant scored 81.2 on AIME 2025 against 65.4 for its predecessor and 76.0 on MMMU-Pro versus 69.2. Developers will access GPT-5.5 Instant through the API as "chat-latest," OpenAI wrote.

Memory sources stop short of full audit

The personalization changes pull on past conversations, uploaded files and connected Gmail accounts. Memory sources, which OpenAI is enabling across all ChatGPT models, let a user tap a button beneath a response to see which prior chats or files informed the answer and to delete or correct any of them. The company acknowledged in its post that the feature "may not show every factor that shaped an answer" and committed to expanding that coverage over time.

For enterprise buyers that constraint matters. Malcolm Harkins, chief trust and security officer at HiddenLayer, told VentureBeat that memory sources "look like a pragmatic middle ground" but are "directionally useful but insufficient on its own," with real value depending on how they integrate with security, governance and audit systems. VentureBeat reported that the partial visibility creates a second context log that can conflict with retrieval and orchestration logs already kept by enterprise stacks.

Rollout, retirement and tone

GPT-5.5 Instant replaces GPT-5.3 Instant as ChatGPT's default for all logged-in users. Paid subscribers can continue selecting GPT-5.3 Instant from the model picker for three months before retirement, OpenAI said. The previous default arrived on March 3 and Tuesday's update is the first version bump in two months. The new personalization features are initially available to Plus and Pro users on the web, with Free, Go, Business and Enterprise access scheduled for the coming weeks, according to Axios.

OpenAI also described tone changes: shorter answers, fewer follow-up questions and a reduction in what its post called "gratuitous emojis." Tom's Guide reported that responses used "roughly 30% fewer words in certain scenarios."

Tuesday's release arrives two weeks after OpenAI shipped GPT-5.5 Thinking and Pro for paid tiers. The Instant line jumped past GPT-5.4 without explanation. Peter Gostev of evaluator Arena told VentureBeat that GPT-5.3-Chat ranked 44th on the Overall Text Arena, 32 places below GPT-5.2-Chat; where GPT-5.5 Instant lands will indicate whether the accuracy work translates to user preference outside OpenAI's own evaluations.

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: [email protected]