SubQ 1M-Preview: First Subquadratic LLM for Long-Context AI

As of May 14, 2026 Subquadratic Labs has shipped SubQ 1M-Preview, the first commercial subquadratic large language model. It carries a 12-million-token context window and matches frontier performance while burning roughly one-fifth the inference compute of conventional transformers. API access opened immediately after the May 5 announcement, aimed squarely at agentic and long-context generative pipelines. Early internal benchmarks already show the model handling extended reasoning chains and multimodal inputs without the usual quadratic blow-up in cost. For anyone who has watched context limits choke detailed storyboards, the numbers feel like a genuine step change rather than incremental marketing.

Traditional transformers pay a quadratic tax on every added token. SubQ sidesteps that scaling wall through an attention mechanism whose compute grows far more linearly. The practical result is that creators can feed entire scene breakdowns, multi-shot scripts or hour-long reference transcripts without the bill exploding. I spent an afternoon stress-testing the preview on a 40,000-token video prompt that would normally trigger timeouts elsewhere. It returned coherent frame-by-frame guidance in one pass. Honestly, that single test made the architecture shift feel less like a research curiosity and more like the new baseline.

On long-context reasoning suites the preview posts scores within 3-4 % of GPT-5.5 while using 78 % less compute at the 1 M token mark. Against the latest Claude Opus variant it trails slightly on creative writing but leads on sustained multimodal coherence once prompts exceed 200 k tokens. Cost per million tokens sits at roughly 22 % of current frontier rates according to the published pricing sheet. Those margins matter when you are iterating on 10-minute video outlines or stitching together 50-image storyboards. The gap is not theoretical; it shows up in real wallet impact after a week of heavy use.

Long, coherent prompts are suddenly cheap enough to treat as first-class creative material rather than an expensive luxury. Directors can now drop full character bibles, lighting references and dialogue tracks into a single call and receive usable shot lists without token gymnastics. The same efficiency gains are already appearing in adjacent creative domains. Advances in multimodal AI are already being applied to adult content creation like in this analysis of Seedance 2.0. My completely unscientific sample of one suggests the real winner will be iterative workflows: generate, review, refine across dozens of passes without watching the meter tick up at the old quadratic rate.