April 2026 broke the AI frontier wide open. GPT-5.5 cracked 60 on the Intelligence Index. Claude Opus 4.7, DeepSeek V4, Kimi K2.6, and MiMo V2.5 Pro all crossed the 50-point threshold in a single month. Five labs pushed the ceiling higher in 30 days than the previous six months combined.
Then May arrived, and the frontier went quiet. But that doesn't mean nothing happened. The action moved sideways — into architecture, efficiency, and open weights. Here's what matters.
The Big Picture
The Intelligence Index ceiling from April (GPT-5.5 at 60.24) has held into late May. No frontier-scale release from Anthropic, Google, Meta, Mistral, or the Chinese labs has broken through yet. But that doesn't mean May is a placeholder month. Three releases are reshaping how we think about what an LLM can be.
SubQ 1M-Preview: Subquadratic Attention Goes Commercial
The most technically interesting release of May isn't a frontier model at all. Subquadratic, a startup with $29M in seed funding, launched SubQ 1M-Preview on May 5 with a simple but radical claim: their model is not a transformer.
Standard transformer attention scales quadratically — double the context length, quadruple the compute cost. That's why "1M context" claims from existing providers come with quiet caveats about quality degradation. SubQ uses sparse, subquadratic attention end to end, and ships with a native 12 million token context window.
At roughly one-fifth the cost of frontier models for long-context tasks, SubQ isn't competing on reasoning benchmarks. It's competing on economics at scale. For document analysis, codebase understanding, and enterprise retrieval tasks where context length actually matters, this could be a genuine category shift.
Why it matters for you: If you're paying premium prices for long-context API calls, SubQ may offer a dramatically cheaper alternative. The architecture is new enough that edge cases will emerge, but the cost-per-token at full context length is compelling.
ZAYA1-8B: Open-Weight MoE That Punches Above Its Weight
Zyphra released ZAYA1-8B on May 6, an open-weight mixture-of-experts model under the Apache 2.0 license. The model activates roughly 760M parameters per token while having 8B total — meaning it runs on hardware that would normally host a 1B dense model, but performs closer to the 3-7B range.
MoE (Mixture of Experts) isn't new, but an open-weight Apache-licensed MoE model with competitive performance at this size is. It's the kind of model that makes local deployment practical on consumer hardware — a laptop with 8GB VRAM can run it comfortably.
Why it matters for you: Self-hosted AI is getting more capable at lower resource requirements. ZAYA1-8B is free, runs locally, and handles most general-purpose tasks well enough for production use.
GPT-5.5 Instant: The New Default
OpenAI didn't just ship GPT-5.5 in April — they followed up with GPT-5.5 Instant on May 5, making it the new ChatGPT default. The model is a distilled version of the full GPT-5.5, optimized for speed while retaining most of its reasoning capability.
This is less a technical story and more a distribution story. When a model becomes the default for hundreds of millions of users, it sets the baseline expectation for what "AI" can do. Every competitor now has to justify why their model is better than what people get for free by default.
Grok 4.3 and Gemini 3.1 Flash Lite
xAI shipped Grok 4.3 on May 6 with incremental improvements over Grok 4. Google released Gemini 3.1 Flash Lite on May 8, positioning it as the fastest option in the Gemini family for latency-sensitive applications.
Neither redefines the landscape, but both illustrate a trend: the mid-tier model market is getting crowded. When five different providers offer models that score within a few points of each other on standard benchmarks, differentiation shifts to latency, cost, context handling, and ecosystem integration.
What This Means for AI Practitioners
Three patterns are emerging in mid-2026:
Architecture diversity is accelerating. The transformer's dominance is being challenged — first by state-space models (Mamba, RWKV), now by subquadratic attention (SubQ). This isn't academic anymore. Commercial products are shipping with non-transformer architectures.
Open weights are closing the gap. ZAYA1-8B, DeepSeek V4, and others show that open-weight models are no longer just "good enough." For many use cases, they're genuinely competitive with proprietary alternatives.
Cost efficiency is the real battleground. When five frontier-class models exist, the race shifts from "who's best" to "who's cheapest at good enough." That's great news for everyone building on top of these models.
Looking Ahead
The labs that sat out May (Anthropic, Meta, Mistral, Google's next flagship) are almost certainly preparing releases for June or July. The frontier will move again. But May's lesson is clear: the most interesting developments in AI aren't always at the top of the leaderboard. Sometimes they're in the architecture underneath.