OpenAI、幻覚が52.5%減少したGPT-5.5 Instantをリリース
OpenAIは、新しいChatGPTのデフォルトモデルとしてGPT-5.5 Instantをリリースし、高リスクの医療・法律系プロンプトにおいてGPT-5.3と比較してハルシネーション(幻覚的な主張)が52.5%減少すると主張しています。フラグされたエラーにおける不正確な主張を37.3%削減し、より簡潔な回答を提供するほか、新しいメモリソース管理機能が含まれます。GPT-5.3 Instantは3ヶ月間引き続き利用可能です。
OpenAI has designated GPT-5.5 Instant as the new default model for ChatGPT, succeeding GPT-5.3 Instant. The company reports that the new model reduces hallucinated claims by 52.5% on high-stakes prompts involving medicine, law, and finance, and decreases inaccurate claims by 37.3% in conversations users previously flagged for errors. The model also features tighter responses with fewer emojis and improved context awareness from sources like Gmail.
The rollout begins immediately for all ChatGPT users, though GPT-5.3 Instant will remain accessible for three months before retirement. Enhanced personalization features and a new 'memory sources' tool—which allows users to view, delete, or correct context used in responses—are launching first for Plus and Pro users on the web, with mobile and other tier support planned.