OpenAI says GPT-5.5 Instant cuts ChatGPT hallucinations by 52.5%

OpenAI said its new default ChatGPT model, GPT-5.5 Instant, cut hallucinated claims by 52.5% compared with GPT-5.3 Instant on high-stakes prompts in medicine, law and finance. The company also said the model reduced inaccurate claims by 37.3% in especially difficult conversations that users had flagged for factual errors.

That matters because OpenAI describes Instant as the daily driver for hundreds of millions of people, a model that now sits in front of students, office workers, patients and anyone treating ChatGPT as a factual answer engine. The company’s pitch is not that hallucinations disappeared, but that even small gains in accuracy can have larger effects when the model is used at scale.

OpenAI said the new system is better at analyzing photo and image uploads, answering STEM questions and deciding when to use web search. It also said GPT-5.5 Instant gives clearer, more concise responses and a more natural conversational tone, while using fewer gratuitous emojis. Those changes are presented as usability improvements, but they sit alongside the harder question that still defines the product: whether the model can be trusted when the stakes are real.

The company’s own framing suggests caution. In September 2025, OpenAI argued that language models hallucinate in part because standard training and evaluation procedures reward guessing over admitting uncertainty. In August 2025, when it introduced GPT-5, OpenAI said the system had made significant progress in reducing hallucinations, especially when reasoning, while also calling hallucinations a fundamental challenge that remains unresolved.

GPT-5 itself was described as a unified system that routes between a fast answer model and a deeper reasoning model, a design meant to balance speed and accuracy. OpenAI’s GPT-5.5 system card, published April 23, 2026, says the model is aimed at complex real-world work, including writing code, researching online, analyzing information, creating documents and spreadsheets, and moving across tools to get things done.

OpenAI has centered the latest upgrade on internal evaluations, not an outside benchmark, which leaves the company asking readers to trust its own measurement of progress. Adam Tauman Kalai, one of the names associated with OpenAI’s work on hallucinations, has become part of a broader company message: factual error is being treated as a measurable engineering problem, but not yet a solved one. For anyone using ChatGPT for medical, legal, financial or classroom answers, that distinction still matters.