Overview
- ChatGPT’s engine, which switched Tuesday to GPT-5.5 Instant, is also live in the API as chat-latest while GPT-5.3 Instant stays for paid users for three months.
- OpenAI’s internal tests report 52.5% fewer hallucinated claims on high‑stakes prompts and 37.3% fewer inaccurate claims on tough, user‑flagged chats.
- Scores climbed on key benchmarks, including AIME 2025 math (81.2 vs. 65.4) and MMMU‑Pro multimodal reasoning (76.0 vs. 69.2), signaling stronger math and visual understanding.
- A new “memory sources” view shows which past chats, files, or connected Gmail shaped a personalized reply, with options to delete or correct those inputs; this rolls out first to Plus and Pro on the web.
- OpenAI labels GPT‑5.5 Instant a High Capability model in cybersecurity and biological domains, pairing the launch with added deployment safeguards detailed in its system card.