OpenAI's GPT-5.5 Instant cuts hallucinations 52.5%, becomes ChatGPT default

OpenAI has updated ChatGPT's default model to GPT-5.5 Instant as of May 5, replacing GPT-5.3 Instant. Internal evaluations show a 52.5% reduction in hallucinated claims across medicine, law, and finance prompts — domains where ground-truth answers are hard to fact-check. User-flagged inaccuracies dropped 37.3%. The model also produces responses that are 30.2% shorter and 29.2% fewer lines, emphasizing substance over filler. For example, a prompt like 'how do I tell my coworker to stop yapping' now yields five tight scripts instead of five tactics plus a 'what not to do' appendix. This release is distinct from the GPT-5.5 flagship (April 23), which is a thinking/agent model with 1M context. GPT-5.5 Instant is positioned as the conversational daily driver for hundreds of millions of users, optimized for retention and speed rather than benchmark scores.

Beyond concision, GPT-5.5 Instant demonstrates improved factuality with self-correction. OpenAI provides a math example: when asked to check work on √(x+7)=x-1, the new model not only detects the error in the user's solution but traces back to find the algebraic mistake (squaring error) and re-derives the correct answer. This 'recovery loop' — revisiting its own prior endorsement when evidence contradicts it — is qualitatively different from models that give confident wrong answers faster. OpenAI frames the model's capabilities around three pillars: factuality with self-correction, concision without losing substance, and reduced follow-up questions. The marginal gains are enormous at scale: trimming 30% of words across billions of conversations compounds to staggering aggregate attention savings for users.