A Startup Says It Cracked AI’s Decade-Old Math Limit — Its LLM Read 12M Tokens for $8
Last Updated on June 22, 2026 by Editorial Team
Author(s): Chew Loong Nian – AI ENGINEER
Originally published on Towards AI.
A Startup Says It Cracked AI's Decade-Old Math Limit — Its LLM Read 12M Tokens for $8
A Miami startup says it ran a long-context job that costs about $2,600 on Anthropic’s top model for $8 on its own LLM, read 12 million tokens in a single pass, and clocked 56x faster than FlashAttention in an independent test. The same week, an AI engineer called the company “either the biggest breakthrough since the Transformer, or it’s AI Theranos.” I spent a day digging through the benchmarks, the skeptics, and the actual math. Here is what holds up and what doesn’t.
The article breaks down Subquadratic (SubQ), a new model architecture aimed at eliminating the transformer’s quadratic attention bottleneck. It explains how SubQ replaces dense attention with dynamically selected Subquadratic Sparse Attention (SSA), keeping only the token pairs that matter so scaling can move toward linear behavior, and it describes what the company offers in private beta (a full-context API, code-loading “SubQ Code,” and long-context “SubQ Search”). It then evaluates the key “receipts” behind the bold claims—especially third-party verified results on RULER and MRCR—while contrasting them with the skeptic case that much of the story may be misrepresented or based on a retrofit rather than a true from-scratch breakthrough. The piece also walks through how to measure the quadratic attention wall yourself and demonstrates an existing linear-scaling alternative (Mamba), using that as a practical baseline for understanding what’s real about “subquadratic” approaches. It concludes with guidance on what to use today (frontier models, RAG, or newer subquadratic options as they mature) and ends with the author’s verdict: SubQ likely isn’t fraud, but also hasn’t fully proven that it has “solved the transformer,” with the economics potentially changing only if the $8 and scale claims hold up under wider scrutiny.
Read the full blog for free on Medium.
Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.
Published via Towards AI
Towards AI Academy
We Build Enterprise-Grade AI. We'll Teach You to Master It Too.
15 engineers. 100,000+ students. Towards AI Academy teaches what actually survives production.
Start free — no commitment:
→ 6-Day Agentic AI Engineering Email Guide — one practical lesson per day
→ Agents Architecture Cheatsheet — 3 years of architecture decisions in 6 pages
Our courses:
→ AI Engineering Certification — 90+ lessons from project selection to deployed product. The most comprehensive practical LLM course out there.
→ Agent Engineering Course — Hands on with production agent architectures, memory, routing, and eval frameworks — built from real enterprise engagements.
→ AI for Work — Understand, evaluate, and apply AI for complex work tasks.
Note: Article content contains the views of the contributing authors and not Towards AI.
Related posts
Recent Posts
