This isn't a flashy launch; no new mode, no jaw-dropping demo. But "small improvement" is a relative term when the upgrade cuts hallucinations by more than half.
What is GPT-5.5 Instant?OpenAI's GPT family ships in tiers. Instant is the everyday model, built for speed and general use; Thinking is the slower, more analytical version for complex problems; and Pro is the heavyweight for maximum-intensity tasks.
GPT-5.5 Instant is the latest update to the tier that most ChatGPT users will interact with, whether they realize it or not.
OpenAI also tested against conversations real users had previously flagged for factual errors. On those, inaccurate claims dropped by 37.3%.
On HealthBench—a benchmark testing AI responses to real medical questions, scored from 0 to 100— GPT-5.5 Instant scores 51.4 points, up from 49.6. On HealthBench Professional, the clinical-use version, it jumps from 32.9 to 38.4 points.
Health questions are among the most common things people ask ChatGPT, which makes getting them right more than a benchmark exercise. These results mean GPT 5.5 Instant increased accuracy by responding correctly 38.4% of the time.
GPT-5.5 Instant also pulls more actively from your past chats, saved files, and connected Gmail account to make answers personally relevant. Now when it does this, it shows you exactly what context it used, and lets you delete or correct it. "You remain in control of what's in your memory," OpenAI wrote. Temporary chats still opt out entirely.
Where it fits—and what it isn'tThe full GPT-5.5 scores 82.7% on Terminal-Bench 2.0, which measures complex command-line task performance. Instant is what the rest of us get, and probably what most of the users will probably be fine working with.
One notable footnote in the system card: GPT-5.5 Instant is the first Instant-tier model OpenAI classifies as "High Capability" in both cybersecurity and biological domains—capable enough to require the same automated safeguards previously reserved for the more powerful Thinking variants. It won't help you hack anything, but OpenAI built guardrails in case someone tries.
