Miami-based startup Subquadratic has published third-party benchmarks for SubQ, its sparse-attention large language model, a month after coming out of stealth with claims that drew immediate skepticism. Evaluation firm Appen measured SubQ at 89.7% on LiveCodeBench and 56 times faster than models using FlashAttention on a baseline speed test. The company also says running Nvidia's RULER 128 long-context benchmark cost $8 on SubQ, compared with $2,600 to run Anthropic's Opus 4.6 through the same test.
The pitch is that SubQ handles up to 12 million tokens of context — roughly 12 times the 1-million-token windows offered by most top models from Google DeepMind, OpenAI, and Anthropic — while approaching their performance on coding and retrieval tasks. Appen's report says SubQ sustained 98% on the needle-in-a-haystack test at both 6 million and 12 million tokens, what the firm called "near-perfect long-context retrieval at scales few models are tested at."
Subquadratic is positioning SubQ as a model purpose-built for coding and searching very large document sets, not a general-purpose replacement for frontier LLMs. The company says tens of thousands of users have signed up for early access, including more than 500 enterprise customers, though a long waitlist means few outsiders have tested the model directly.
Key facts
- 01Subquadratic says SubQ ran Nvidia's RULER 128 test for $8, versus $2,600 to run Anthropic's Claude Opus 4.6 through the same benchmark.
- 02Appen measured SubQ at 89.7% on LiveCodeBench and 56x faster than FlashAttention models in a baseline speed test.
- 03SubQ's context window stretches to 12 million tokens, roughly 12x the 1-million-token windows on most top models, with 98% scores on needle-in-a-haystack at 6M and 12M tokens.
- 04The Miami startup says tens of thousands of users and 500+ enterprise customers have signed up for early access, with a long waitlist.
- 05SubQ was bootstrapped from the weights of the Chinese open-source model Qwen rather than trained from scratch.
Cofounder and CEO Justin Dangel argues the architecture is the real story. The 2017 Google paper that defined the modern era, "Attention Is All You Need," introduced the transformer and its dense-attention mechanism. Dense attention encodes each token as a number and multiplies it against every other token in the input — a 10,000-word passage triggers nearly 50 million individual multiplications, and the cost scales quadratically with length.
Sparse attention, which SubQ uses, only multiplies a chosen subset of those token pairs. The approach has been tried for years, by labs inside and outside the major frontier shops, and has consistently failed to match dense attention on broad benchmarks. Subquadratic's claim is that it has finally closed the gap.
Cofounder and CTO Alex Whedon says the difference is how SubQ picks which token pairs matter. Earlier sparse-attention work relied on fixed patterns; SubQ selects dynamically on a per-input basis. Whedon would not detail the selection mechanism, calling it the "secret sauce."
Appen's Jeanine Sinanan-Singh, who directs the firm's generative AI research, said the results validated the architecture and called SubQ's coding performance "frontier-level." In a live demo, Whedon had SubQ reason across 400 documents in seconds; Perplexity, given the same task, failed to load the documents.
The skepticism is not unwarranted. When Subquadratic first announced SubQ, it published only self-reported scores, prompting AI engineer Dan McAteer to summarize the reaction on X: "SubQ is either the biggest breakthrough since the Transformer ... or it's AI Theranos." Whedon told MIT Technology Review the company should have released third-party benchmarks alongside the initial announcement.
Two caveats sit on top of the new numbers. Benchmarks, even credible third-party ones, do not substitute for sustained use across messy real-world tasks, and Appen's evaluation covers a handful of standard tests. Subquadratic also bootstrapped SubQ from the weights of the Chinese open-source model Qwen rather than training from scratch — a common shortcut, but one that complicates comparisons against models built end-to-end on a new architecture. Until more developers get hands-on access, the question of how SubQ behaves at the edges remains open.
If the architecture holds up under wider use, the cost gap is the part worth watching. A 325x difference in benchmark cost — $8 versus $2,600 on RULER 128 — is the kind of delta that reshapes which workloads economically justify long-context reasoning at all. Document analysis over hundreds of files, whole-codebase reasoning, and large-scale retrieval pipelines are tasks that today's frontier models can do but few customers can afford to run at volume. A model that closes most of the performance gap while cutting inference cost by orders of magnitude does not have to win on every benchmark to take meaningful share of the workloads where context length is the constraint. Whether SubQ is that model, or whether the next round of independent testing exposes cracks, is what the next few months of broader access will settle.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
