For eighteen months, the frontier AI race looked like a horsepower competition: bigger models, higher benchmark scores, more parameters. May 2026 broke that pattern. According to model trackers at WhatLLM.org, the Intelligence Index ceiling held flat for the first time in a year — no new model topped the scorecard. What moved instead was architecture. The headline act is SubQ 1M-Preview, from a San Francisco startup called Subquadratic, which shipped its first commercially available model built entirely on sparse, subquadratic attention — dispensing with the transformer's fundamental O(n²) computational bottleneck.
The practical result is a native 12-million-token context window and a claimed 52x speed advantage on long-sequence attention tasks, at roughly one-fifth the inference cost of frontier rivals. Independent verification is still pending, but the architecture claims are drawing serious attention from researchers who have spent years arguing that standard attention would eventually hit a wall. The month's second surprise came from Zyphra, an open-source lab that trained ZAYA1-8B entirely on AMD Instinct hardware — marking the first reasoning-focused open release built on AMD silicon rather than NVIDIA. With only 760 million parameters active per token despite an 8-billion total parameter count, ZAYA1-8B exemplifies what researchers are calling "intelligence density": getting more reasoning per active compute than brute-force scaling allows.
Meanwhile, OpenAI quietly made GPT-5.5 Instant the default ChatGPT model on May 5, emphasizing fewer hallucinations in law, medicine, and finance — a tacit acknowledgment that trust, not raw capability, is now the product. The efficiency turn in AI model development is real and compounding. The next six months will test whether subquadratic architectures can hold up under production load or whether the benchmark gap with standard transformers is wider than advertised. If SubQ's claims survive independent verification, the cost economics of long-context inference will shift dramatically — unlocking use cases that are currently too expensive to run at scale.