GPT-5.5 Instant Launch Cuts Hallucinations by 52.5% and Eliminates Fluff

Math Reasoning Improvement

OpenAI’s side‑by‑side test shows that when solving an algebra equation, GPT‑5.3 Instant abandons the task after the initial solution fails, whereas GPT‑5.5 Instant backtracks, checks the algebraic error and derives the correct answer, demonstrating useful self‑correction.

Hallucination Reduction 52.5%

Internal tests in high‑risk domains such as medical and legal fields report that the probability of fabricating facts drops by more than half. User‑marked incorrect answers decrease by 37.3%.

On academic benchmarks, GPQA doctoral‑level science accuracy rises from 78.5% to 85.6%, and AIME math competition scores improve by 15.8 points.

Goodbye to Fluff

In a user test asking “how to boil an egg”, answer length shortens by an average of 42%. Excessive formatting, redundant rhetorical questions and emojis are noticeably reduced.

For workplace communication suggestions, word count falls 30.2% and line count 29.2%, with the model delivering concise scripts instead of lengthy prose.

Memory Becomes a Practical Feature

When asked “recommend a new tea‑drink shop”, the model proactively pulls Gmail reservation records and past chat taste preferences. The new “memory source” feature makes each personalized response traceable to its data source and allows users to delete referenced content.

Gray‑Rollout Strategy

Free users can use the new model immediately; the API alias is chat-latest.

Paid users may manually switch back to GPT‑5.3 Instant, with the option retained for three months.

The personalized memory feature is initially available to Plus/Pro web users.

A developer noted that the smartest design choice is preserving the “Instant” positioning—focusing on being “good enough and handy” for everyday scenarios rather than trying to fully surpass GPT‑5.5 Thinking. When most users do not even notice the model change, the experience win is achieved.