GPT-5.5 tops AA index at every tier

The real story here is less “new benchmark king” than “OpenAI is squeezing more score out of less thinking budget.” If these numbers hold up in production, medium may become the default sweet spot while xhigh stays a niche bragging-rights mode. Artificial Analysis lists a tight 60/59/57 spread across xhigh, high, and medium, which suggests diminishing returns as reasoning effort increases. Medium at 57 is the eye-catcher because it implies GPT-5.5 can hit frontier-level benchmark performance without forcing developers into the slowest, most expensive setting. OpenAI’s launch post also leans hard on token efficiency, arguing GPT-5.5 reaches higher-quality outputs with fewer tokens and fewer retries; that matters more for real workloads than a single headline score bump. Artificial Analysis flags at least some GPT-5.5 benchmark results as lab-claimed and not yet independently verified, so developers should treat the leaderboard as directional until more third-party testing lands. Reddit’s reaction is already splitting along the expected line: impressive efficiency gains on one side, “benchmaxxing” skepticism on the other.