OpenAI Claims GPT-5.5 Instant Cuts Hallucinations by 52%

Key Takeaways

• OpenAI claims GPT-5.5 Instant produces 52.5% fewer hallucinated claims than GPT-5.3 Instant on medical, legal, and financial prompts
• The new model also reduced inaccurate claims by 37.3% on conversations users had flagged for errors
• GPT-5.3 Instant will remain available for three months before being retired

OpenAI announced Tuesday that its new default ChatGPT model hallucinates less than its predecessor. The company says GPT-5.5 Instant produces 52.5% fewer made-up claims compared to GPT-5.3 Instant when handling prompts about medicine, law, and finance.

The numbers come from OpenAI's internal evaluations, not independent testing. Still, the specific percentage marks a departure from the vague "improved accuracy" claims that typically accompany AI model updates.

What OpenAI Is Claiming

Beyond the headline number, OpenAI says GPT-5.5 Instant also reduced inaccurate claims by 37.3% on "especially challenging conversations users had flagged for factual errors." This suggests the company is using real user complaints to benchmark accuracy, not just synthetic test cases.

The company describes the new model as "more capable across everyday tasks." That includes better analysis of uploaded images and smarter decisions about when to search the web for answers instead of relying on training data.

OpenAI also addressed a complaint that has annoyed many ChatGPT users: excessive emoji. The company says GPT-5.5 Instant will avoid "gratuitous emojis" and deliver responses that are "tighter and more to-the-point."

Enhanced Personalization and Memory

GPT-5.5 Instant can now pull more context from previous chats and connected services like Gmail. This means responses can be tailored to your history and preferences. Google has been building similar features into Gemini, so personalization is becoming a competitive battleground.

A new "memory sources" feature will show users exactly what context ChatGPT used to personalize a response. You can delete or correct this information if it's wrong or outdated. This addresses a common concern: when AI remembers things about you, you should be able to see and control what it knows.

The enhanced personalization features roll out first to Plus and Pro subscribers on the web, with mobile apps coming later. Free, Go, Business, and Enterprise users will get access "soon," though OpenAI didn't specify a date.

Rollout and Transition Period

OpenAI starts rolling out GPT-5.5 Instant on Tuesday to all ChatGPT users. The company is keeping GPT-5.3 Instant available as an option for three months before retiring it.

The transition period acknowledges a real pattern: users sometimes prefer older models. When OpenAI has retired previous versions, some users publicly mourned the loss. Three months gives people time to adjust or find workarounds if they prefer the older model's behavior.

The memory sources feature is rolling out now to ChatGPT consumer plans on the web, with mobile support coming soon.

The Hallucination Problem

Hallucinations remain one of the biggest barriers to AI adoption in professional settings. A model that confidently states false medical information or cites non-existent legal precedents isn't just wrong. It's dangerous.

OpenAI's specific claims about improvement percentages are notable, but they come with caveats. "Internal evaluations" means we're taking OpenAI's word for it. Independent benchmarks would carry more weight, especially for organizations considering AI deployment in regulated industries.

The 52.5% reduction sounds impressive, but context matters. If the previous model hallucinated on 20% of high-stakes prompts, a 52.5% reduction means GPT-5.5 Instant still hallucinates on roughly 9.5% of them. That's better, but still far from reliable for unsupervised use in medicine, law, or finance.

Frequently Asked Questions

When does GPT-5.5 Instant become the default ChatGPT model?

OpenAI begins rolling it out on Tuesday, May 5, 2026, to all ChatGPT users.

How much less does GPT-5.5 Instant hallucinate compared to the previous model?

OpenAI claims 52.5% fewer hallucinated claims on high-stakes prompts covering medicine, law, and finance, and 37.3% fewer inaccurate claims on conversations users flagged for errors.

Can I still use GPT-5.3 Instant after the update?

Yes, GPT-5.3 Instant will remain available as an option for three months before OpenAI retires it.

What is the new memory sources feature in ChatGPT?

Memory sources shows you what context ChatGPT used to personalize a response, letting you delete or correct information if needed.

Who gets the enhanced personalization features first?

Plus and Pro subscribers on the web get access first, with mobile apps and other subscription tiers coming later.

New 'Memory Sources' Feature and Rollout Details

The new article introduces the 'memory sources' feature, which shows users exactly which personal context or files informed a specific AI response. It also provides specific rollout details for Plus and Pro subscribers and mentions that the model is available via the API as 'chat-latest'.

Enhanced Personalization and User Memory Controls

The article introduces 'memory sources,' a new feature that gives users visibility and control over the data used to personalize responses, including the ability to delete or update specific memories. It also details performance improvements for searching uploaded files and connected Gmail accounts, along with a phased rollout starting with Plus and Pro subscribers.

Improved STEM performance and technical bug reports

The new article reveals that GPT-5.5 Instant is more capable at STEM tasks, image analysis, and deciding when to use web search. It also highlights a specific algebra correction example and mentions a potential 'Goblin' issue previously identified in the model.