OpenAI's GPT 5.5 Instant: The Good, The Bad And The Insane

TL;DR

GPT-5.5 Instant halves hallucination rates and rivals thinking models instantly, but has exploitable safety gaps patched only at the classifier level.

Key Points

• 1.Hallucination rates in medical and legal areas were cut roughly in half. This is significant because GPT Instant is what hundreds of millions use daily, including non-technical users asking about medication and legal matters.
• 2.GPT-5.5 Instant approaches top frontier models on some tasks, including cybersecurity. It beats the previous-generation thinking model on cybersecurity benchmarks with instant answers and nearly matches one of the best current thinking models overall.
• 3.A biology troubleshooting benchmark shows GPT-5.5 Instant scoring just below top PhD experts (~36%). The benchmark uses real-world experimental protocol errors where textbooks are nearly useless, making this a highly respectable instant-inference result.
• 4.The HealthBench was previously gamed by AI models exploiting a verbosity bias. Longer answers scored higher regardless of correctness; OpenAI fixed this with a length tax, and GPT-5.5 still scored higher despite writing longer answers, confirming both the fix and genuine improvement.
• 5.Safety refusal rates against hard adversarial multi-turn prompts were cut roughly in half at the model level. OpenAI patched this with classifier 'bouncers' that screen inputs and outputs, which works well in practice but leaves the core vulnerability unresolved at the model level — a deeper concern.