OpenAI cuts ChatGPT hallucinations by over 50% with new model
What's the story
OpenAI has launched a new foundation model, dubbed GPT-5.5 Instant, as the default for its ChatGPT service. The company claims the new model "produced 52.5% fewer hallucinated claims than GPT‑5.3 Instant on high-stakes prompts covering areas like medicine, law, and finance." The release of GPT-5.5 Instant comes after last month's introduction of the latest GPT-5.5 model with improvements in coding and knowledge work capabilities.
Performance metrics
New model scores better on various benchmarks
The new GPT-5.5 Instant model has scored 81.2 in the AIME 2025 math test, a huge improvement over the previous version's score of 65.4. It also surpassed its predecessor on the MMMU-Pro multimodal reasoning benchmark with a score of 76.0 against 69.2 for the older model. These improvements highlight OpenAI's commitment to enhancing its AI technology and providing users with better performance across various tasks and applications.
User experience
GPT-5.5 Instant can manage context better
The release of GPT-5.5 Instant also focuses on improving context management. The model can use its search tool to refer back to past conversations, files, and Gmail for more personalized responses. This feature will be available for Plus and Pro users on the web, with a mobile rollout planned soon. OpenAI plans to extend access to this feature for Free, Go Business, and enterprise users in the coming weeks.
Developer access
Other notable changes in ChatGPT
Along with the new model, ChatGPT will also show memory sources across all models. This is to help users understand where their answers were generated from. Users can delete or correct outdated sources if needed. The GPT-5.5 model will be available through API as 'chat-latest,' with 5.3 remaining an option for paid users for three months.