GPT-5.5 Instant becomes ChatGPT default; hallucinations drop

OpenAI on Tuesday replaced the default model in ChatGPT with GPT-5.5 Instant. The company said the model is available to all users for free and produced 52.5% fewer hallucinated claims than GPT-5.3 Instant on selected high‑stakes medical, legal and financial prompts in internal tests.

OpenAI describes Instant as the everyday, fast tier for general use, while Thinking and Pro are for heavier analytical or high‑intensity tasks. The change does not accompany a new interface or public demonstration; the company released performance figures and deployment details instead.

In its internal evaluations, OpenAI reported a 37.3% reduction in incorrect claims when GPT-5.5 Instant was tested against conversations that users had previously flagged for factual errors. On HealthBench, which scores AI responses to real medical questions on a 0–100 scale, GPT-5.5 Instant scored 51.4, up from 49.6 for the prior Instant model. On HealthBench Professional, the clinical-use version, the model improved from 32.9 to 38.4 points.

GPT-5.5 Instant can draw more actively from a user’s past chats, uploaded files and a connected Gmail account to produce context‑aware answers. When the model consults stored context, ChatGPT will display the specific items it used and give users the option to delete or correct them. OpenAI wrote, “You remain in control of what’s in your memory.” Temporary chats remain excluded from memory features.

OpenAI classifies GPT-5.5 Instant as the first Instant‑tier model with “High Capability” in both cybersecurity and biological domains. The company said extra automated safeguards are in place at deployment to limit misuse and that the model will not provide assistance for hacking.

The rollout is phased by feature. Enhanced personalization via Gmail will be available first to Plus and Pro subscribers on the web, with Free, Go, Business and Enterprise accounts to follow in the coming weeks. Paid users who prefer GPT-5.3 Instant will have three months before that version is retired; GPT-5.3 launched in March.

OpenAI also published performance data for the broader GPT-5.5 family. The full GPT-5.5 scored 82.7% on Terminal‑Bench 2.0, a test of complex command‑line tasks. Instant remains the default option tuned for speed and routine tasks such as drafting emails, meal planning and basic queries.

OpenAI’s published figures are based on internal testing. The company said it will continue to roll out features and safety controls incrementally across its user base.