OpenAI Makes GPT-5.5 Instant The Default ChatGPT Model

OpenAI is putting a higher capability Instant model in front of ChatGPT users while promising fewer hallucinations, tighter answers, and clearer memory controls.

SAN FRANCISCO, California - OpenAI said GPT-5.5 Instant is replacing GPT-5.3 Instant as the default ChatGPT model for all users, making the company's newest everyday model the main interface for consumer ChatGPT and the API route called chat-latest.

The practical change for users is simple: the default ChatGPT experience should become more concise, more accurate, and more personalized, according to OpenAI's May 5 announcement. The policy question is harder. OpenAI's deployment safety hub says GPT-5.5 Instant is the first Instant model the company is treating as High capability in both Cybersecurity and Biological and Chemical Preparedness.

Why It Matters

Instant models are the high volume lane of ChatGPT. They are optimized for the default path most people use when they ask questions, upload images, plan work, or ask for help with documents. That makes this launch different from a niche model card or a benchmark update aimed mainly at developers.

OpenAI says GPT-5.5 Instant is now the default model for all ChatGPT users and will be available through the API as chat-latest. Paid users can keep GPT-5.3 Instant for three months through model configuration settings before OpenAI retires it, according to the company's announcement.

The mechanism is not just a faster answer box. OpenAI is changing the model behind the default ChatGPT surface, giving it broader access to personalization context when users have enabled it, and letting the model decide when web search or user history can improve an answer.

The Story So Far

OpenAI has used the Instant label for models designed to answer quickly in the standard ChatGPT flow, while reserving more expensive or slower reasoning models for tasks that need extended deliberation. In its GPT-5.5 Instant announcement, OpenAI described the new model as the daily driver for hundreds of millions of people.

The company said GPT-5.5 Instant gives stronger and tighter answers across subject areas, improves photo and image upload analysis, answers STEM questions better, and makes better decisions about when to use web search. Those are product claims, not architecture disclosures. OpenAI did not publish enough detail to show whether the gains come from new training methods, post training, routing, system instructions, serving changes, or a mix of those factors.

That distinction matters for enterprises and regulators. A model can feel better because its underlying weights changed, because its product wrapper changed, because it searches the web more effectively, or because it uses more personal context. OpenAI's public material points to a deployment and product update more than a disclosed architecture change.

What's Happening Now

OpenAI said GPT-5.5 Instant produced 52.5% fewer hallucinated claims than GPT-5.3 Instant in internal evaluations on high stakes prompts covering medicine, law, and finance. It also said inaccurate claims fell 37.3% on especially challenging conversations users had flagged for factual errors.

Those numbers are meaningful, but OpenAI's own description limits how broadly they should be read. The cited prompts were selected to test high stakes and difficult cases. They do not measure the average factual error rate across normal ChatGPT traffic.

OpenAI also said GPT-5.5 Instant can use past chats, uploaded files, and connected Gmail when available. The company said the model decides when personalization would improve an answer, searches past conversations faster, and can use memory sources that show some of the context used to personalize a response.

The company said users can delete chats they do not want cited, delete or change saved memories in settings, or use temporary chats that do not use or update memory. OpenAI also said memory sources may not show every factor that shaped an answer, because the feature can show some relevant past chats rather than every past chat searched or referenced.

The Safety Card

OpenAI's system card makes the launch more significant than a routine model refresh. The card says GPT-5.5 Instant is the first Instant model OpenAI is treating as High capability in Cybersecurity and Biological and Chemical Preparedness.

"This is the first Instant model that we are treating as High capability in our Cybersecurity and Biological & Chemical Preparedness categories, and implementing appropriate safeguards." - OpenAI, GPT-5.5 Instant System Card

That statement does not mean OpenAI says the public ChatGPT product will provide cyber abuse instructions or biological assistance. It means OpenAI's internal preparedness framework places the underlying model in a higher capability category for those domains, and the company says it is applying safeguards around deployment.

OpenAI said its disallowed content evaluations run on the base model without system level safeguards, so the company can test the model's underlying behavior against its safety bar. It also said it continues monitoring after launch to evaluate online performance and adjust safeguards.

"Our evals are run on the base model, without system-level safeguards, to ensure the model’s underlying behavior meets our safety bar." - OpenAI, GPT-5.5 Instant System Card

The system card says GPT-5.5 Instant is comparable to GPT-5.3 Instant across disallowed categories except gore and disallowed sexual content, where OpenAI applies additional system level mitigations. OpenAI said users it believes may be under 18 receive additional age appropriate protections that further restrict sexual content and exposure to gore.

What Changed In Health And Factuality

OpenAI's safety card reports gains on several health benchmarks. GPT-5.5 Instant improved by 1.8 points on HealthBench, 2.7 points on HealthBench Hard, and 5.5 points on HealthBench Professional compared with GPT-5.3 Instant, while HealthBench Consensus was effectively flat at plus 0.03, according to the system card.

The company also disclosed a scoring caveat. OpenAI said HealthBench and HealthBench Professional can reward longer answers, so it now reports adjusted scores that penalize unnecessary length. That makes the health figures more useful than a raw verbosity contest, but it also means readers should compare the results within OpenAI's stated methodology rather than as a universal measure of medical safety.

"In internal evaluations, GPT-5.5 Instant produced 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts covering areas like medicine, law, and finance." - OpenAI, GPT-5.5 Instant announcement

For users, the most visible improvement may be shorter answers that stay useful. OpenAI said the model reduces verbosity and overformatting, asks fewer unnecessary follow up questions, and avoids clutter such as gratuitous emojis.

The Enterprise And Policy View

Enterprise buyers will likely focus on three practical questions. First, whether the default model's lower hallucination rate appears in their own workflows. Second, whether memory and Gmail context controls meet internal privacy and compliance requirements. Third, whether OpenAI's High capability classification changes the risk review for teams using ChatGPT or chat-latest in regulated environments.

NIST's AI Risk Management Framework gives those buyers a common language. NIST says the framework is voluntary and is meant to improve how organizations incorporate trustworthiness into the design, development, use, and evaluation of AI products, services, and systems.

"The NIST AI Risk Management Framework (AI RMF) is intended for voluntary use and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems." - National Institute of Standards and Technology

NIST also released a Generative AI Profile in 2024 and a 2026 concept note for trustworthy AI in critical infrastructure. Those documents do not regulate OpenAI's launch directly. They do show how U.S. agencies frame the risk management burden when AI systems move from lab tests into operational products.

Other Perspectives

OpenAI's view is that default ChatGPT should become more useful without asking users to choose among model names. The company said GPT-5.5 Instant gives clearer answers, better factuality on hard prompts, and more useful personalization while adding memory source controls.

A privacy focused view starts with the same feature and reaches a different priority. Personalization from past chats, files, and connected Gmail can reduce repeated context, but it also makes user control, source visibility, deletion, and regional availability central to trust. OpenAI acknowledged that memory sources may not show every factor that shaped an answer.

A national security view centers on the safety classification. A U.S. lab shipping higher capability AI through a mainstream consumer product supports American AI leadership. It also gives regulators and security officials a sharper question: how should frontier capability thresholds apply when the model is not a research preview, but the default chat interface for everyday users?

By The Numbers

• 52.5%: OpenAI's reported reduction in hallucinated claims versus GPT-5.3 Instant on high stakes prompts covering medicine, law, and finance.
• 37.3%: OpenAI's reported reduction in inaccurate claims on difficult conversations users had flagged for factual errors.
• 3 months: How long paid users can keep GPT-5.3 Instant before OpenAI retires it.
• 5.5 points: OpenAI's reported GPT-5.5 Instant gain on HealthBench Professional versus GPT-5.3 Instant.
• 2 categories: Cybersecurity and Biological and Chemical Preparedness categories where OpenAI treats GPT-5.5 Instant as High capability.

What People Are Saying

"GPT-5.5 Instant is rolling out starting today to all ChatGPT users, replacing GPT-5.3 Instant as the default model, and in the API as chat-latest." - OpenAI, GPT-5.5 Instant announcement

"Instant is now more effective at using context from past chats, files, and Gmail, if you have it connected, so answers feel more personally relevant while keeping you in control." - OpenAI, GPT-5.5 Instant announcement

"We find that GPT-5.5 Instant is comparable (i.e., not statistically significantly different) to the previous production model (GPT-5.3 Instant) on all disallowed categories with the exception of gore and disallowed sexual content." - OpenAI, GPT-5.5 Instant System Card

"Released on January 26, 2023, the Framework was developed through a consensus-driven, open, transparent, and collaborative process." - National Institute of Standards and Technology

The Big Picture

GPT-5.5 Instant puts OpenAI's default ChatGPT path at the center of the next AI safety debate. The launch pairs consumer facing improvements with a system card that says the underlying model has crossed higher capability thresholds in sensitive domains.

The next test is operational, not rhetorical. Users will see whether the model is actually more concise and accurate in daily use. Enterprises will test whether memory controls and factuality claims hold up in their own data. Policymakers will watch how OpenAI's safeguards perform now that a higher capability Instant model is the default door into ChatGPT.