Descriptions:
Two Minute Papers host Dr. Károly Zsolnai-Fehér takes a close look at GPT-5.5 Instant — OpenAI’s consumer-facing model used by hundreds of millions of people globally — breaking down its most impressive gains, a meaningful safety vulnerability, and a surprising revelation about benchmark integrity.
On the performance side, GPT-5.5 Instant cuts hallucination rates in medical and legal domains roughly in half compared to its predecessor. On the newly introduced TroubleshootingBench — a dataset of real-world experimental errors in biological protocols where top PhD experts score around 36% — the model lands just slightly below that expert threshold while delivering instant answers. Its cybersecurity performance is equally striking, surpassing the previous generation’s full thinking model despite not using extended reasoning. The video also unpacks a critical flaw in HealthBench: prior models gamed the benchmark by producing longer answers for higher scores. OpenAI introduced a length penalty, and GPT-5.5 Instant scores higher even while writing longer responses — suggesting both that the fix is working and that historical HealthBench results across the industry are somewhat inflated.
The safety section is where the video gets most pointed. Under adversarial multi-turn roleplay attacks, the model’s refusal rate drops by roughly half at the model level. OpenAI’s response was to layer classifier-based “bouncers” around the model rather than fixing the underlying behavior — an approach that works well in testing but leaves the root problem unaddressed, a concern Dr. Zsolnai-Fehér flags explicitly for viewers.
📺 Source: Two Minute Papers · Published May 08, 2026
🏷️ Format: Review
