OpenAI cuts ChatGPT hallucinations by 52% with new GPT-5.5 Instant model

OpenAI has introduced an updated default model for ChatGPT, GPT-5.5 Instant, which is markedly more honest and makes things up far less often.

The numbers

Hallucinations, where a model invents false information, have long frustrated ChatGPT users. Based on internal tests, OpenAI says GPT-5.5 Instant produced 52.5% fewer fabricated facts than the previous Instant model for GPT-5.3, especially on high-risk questions in medicine, law and finance. In difficult conversations that users themselves flagged as containing errors, the new model cut inaccurate claims by 37.3%. That is significant progress in areas where a mistake costs money or health.

Where it helps most

The improvements show up where errors matter most:

• Medicine and diagnostics
• Legal advice and interpretation of laws
• Financial planning and investing
• Complex technical questions
• Fact-checking and information verification

Where errors can still happen

OpenAI does not hide that this is not a cure-all. Hallucinations have decreased but have not disappeared entirely. The model can still make mistakes on very recent facts that are absent from its training data, and on highly specialized questions that require rare expertise.

What it means

Making GPT-5.5 Instant the default signals that OpenAI is working seriously on reliability. For users who rely on ChatGPT for critical information, the improvement is noticeable. Healthy skepticism is still necessary, though: verify facts before important decisions and do not trust answers blindly.