The transformer’s biggest flaw just got fixed. Every LLM you’ve ever used compares every token to every other token. That’s O(n²) scaling. Double the context, quadruple the compute. It’s why long-context models are slow and expensive, and why they degrade as context grows. SubQ doesn’t do that. It’s the first frontier model built on a fully sub-quadratic sparse-attention architecture. It finds the token relationships that actually matter and skips the rest. At 12 million tokens, this cuts attention compute by almost 1,000x. The numbers: → 12M token context (others advertise 1M, accuracy collapses past 200k) → 98% accuracy at 12M tokens → 52x faster than FlashAttention at 1M → Less than 5% the cost of Opus 4.7 → 97.1% on RULER 128K vs Opus 4.7 at 97.2% → 83 on MRCR v2 vs GPT-5.4 at 39 Well done

t.co/bdaeTP7pCz