GPT-5.5 Instant Launches: 52.5% Reduction in Hallucinations, Less Verbose Responses

By AI Engineering | May 6, 2026 | Beijing

OpenAI has quietly completed a silent upgrade to ChatGPT’s default model, with GPT-5.5 Instant fully replacing GPT-5.3 Instant. This model, used by hundreds of millions of users as their “daily driver,” brings improvements that directly address practical pain points.

Enhanced Mathematical Reasoning Capabilities

A key comparison showcased on the official website demonstrates GPT-5.5 Instant’s superior problem-solving approach. When solving algebraic equations, GPT-5.3 Instant would give up after discovering an initial solution didn’t work, while GPT-5.5 Instant can backtrack to check for algebraic errors and re-derive the correct solution. This self-correcting ability proves highly practical for everyday use.

52.5% Reduction in Hallucinations

Internal testing reveals that in high-stakes fields like healthcare and law, the probability of fabricating facts has been reduced by more than half. User-flagged incorrect responses have decreased by 37.3%. Academic benchmarks also show improvement: accuracy on GPQA doctoral-level science questions increased from 78.5% to 85.6%, while AIME math competition scores jumped by 15.8 percentage points.

Less Verbose Responses

User testing shows that for simple questions like “how to boil an egg,” the new version’s responses are on average 42% shorter. Excessive formatting, redundant rhetorical questions, and emoji usage have been significantly reduced.

In workplace communication advice comparisons, GPT-5.5 Instant uses 30.2% fewer words and 29.2% fewer lines, providing practical scripts instead of lengthy explanations.

Memory Becomes a Practical Feature

When users ask for “new tea shop recommendations,” the model now actively references Gmail reservation records and taste preferences mentioned in previous conversations. The new “memory source” feature makes every personalized response traceable to its data source, with the ability to delete referenced content at any time.

Gradual Rollout Strategy

• Free users get immediate access to the new model via the API alias chat-latest
• Paid users can manually switch back to GPT-5.3 Instant for three months
• Personalized memory features are initially available to Plus/Pro web users

Developers note that the smartest aspect of this upgrade is maintaining the “Instant” positioning — it doesn’t aim to completely surpass GPT-5.5 Thinking, but rather to be “sufficient and convenient” for everyday scenarios. When most users don’t even notice the model switch, that might be the true victory in user experience.

Join our community discussion by replying “join group” to our official account.