3 min read
OpenAI just swapped out the engine inside ChatGPT. Starting today, GPT-5.5 Instant replaces GPT-5.3 Instant as the default model used by hundreds of millions of people who open ChatGPT every day.
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.
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.
According to OpenAI, the new model produced fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts in medicine, law, and finance. Hallucinations have been ChatGPT's most persistent flaw since the beginning.
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.
When Decrypt covered the GPT-5.5 family launch two weeks ago, the story was agentic coding and terminal workflows. GPT-5.5 Instant is a different animal—it handles more “basic” stuff like your meal plans and email drafts, not autonomous multi-step coding pipelines. Don’t ask us about GPT 5.4 Instant, though. It’s probably chilling next to the O2 model that never existed.
The 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.
The previous default, GPT-5.3 Instant, launched in March with promises of fewer preachy refusals and better accuracy. GPT-5.5 Instant continues that trajectory. Paid subscribers who prefer the old version have three months before GPT-5.3 Instant is retired. Enhanced personalization via Gmail rolls out first to Plus and Pro users on the web, with Free, Go, Business, and Enterprise to follow in the coming weeks.
Decrypt-a-cookie
This website or its third-party tools use cookies. Cookie policy By clicking the accept button, you agree to the use of cookies.