Real company.
Unverified claims.

Before any benchmark argument, the threshold question is whether the company exists at all. Subquadratic clears that bar comfortably. It filed an SEC Form D in February 2026, incorporated as Subquadratic AI, headquartered in Miami. It has a CEO with a twenty-five-year operating record, a CTO with a verifiable engineering background, a live careers page on Rippling, and a research presence at AAAI and ICLR 2026. The company is not a shell.

The uncomfortable parts are not the existence questions. They are the composition questions. Most of the founding team is not ML-research-credentialed in the usual sense. The lead investors lean consumer and marketplace, not foundation-model. The eleven PhD researchers reportedly on staff are unnamed. Each detail in isolation is innocent. The accumulation is what registers.

The honest split is two columns. Real receipts on the left. Real asterisks on the right. Both columns are accurate. Neither cancels the other.

The cluster of numbers above is what triggered the "AI Theranos" reaction on X. Each one sounds extraordinary on first read. Each one carries a different status under inspection. The third-party verifier confirmed two production scores. Everything else, including the most quoted figures, sits in the self-reported, single-run column. That distinction matters more than any individual headline.

The cleanest read of the benchmark sheet is a side-by-side bar chart of MRCR v2, the multi-reference long-context retrieval benchmark that frontier models currently disagree on most. SubQ's production score on MRCR v2 (third-party verified) is 65.9. Their research-config score on MRCR v2 (self-reported) is 83.0. The delta between those two numbers (seventeen points on the same benchmark for the same family of models) is itself larger than the gap between most frontier models' end-to-end results.

The prior is not that the production number is wrong. The prior is that any seventeen-point self-vs-third-party gap on a reproducible long-context benchmark gets independently reproduced before it is treated as load bearing. No reproduction yet exists.

Frontier models on MRCR v2 cluster in a wide band. GPT-5.5 leads the named comparators at 74. Claude Opus 4.7 sits at 32.2 and Gemini 3.1 Pro at 26.3, in part because MRCR v2 is unusually hostile to general-purpose long-context behavior. The headline result for SubQ's production model (65.9) is below the top of that band but well above the rest, which is plausible for a long-context-specialized architecture. The headline result for SubQ's research config (83) is nine points above the strongest frontier reference, on a benchmark where the differences between leading models are usually measured in single digits.

The four other widely-cited scores carry similar caveats:

Sparse attention is not new. Longformer, BigBird, Reformer, and DeepSeek Sparse Attention all narrow the cost of attention by letting each query look at only a subset of keys. The hard part is not "be sparse." The hard part is choosing which keys to keep cheaply enough that the choosing step is itself subquadratic, and keeping quality competitive with dense at frontier scale, and not smuggling a quadratic layer back in to make the whole thing work. Three constraints. Most prior work has hit two.

SubQ claims all three. It calls its mechanism "Subquadratic Selective Attention," pitched as content-dependent and pure (no hybrid quadratic layers). If true, it would be genuinely novel and structurally different from the prior families. If not, it is a marketing reframing of well-studied techniques. Without an arXiv preprint or open weights, the public has the marketing but not the mechanism.

The reason this matters is that the field has a long ledger of claims that did not survive scrutiny. Mamba and RWKV are linear but underperform dense at frontier scale. Mamba-attention hybrids are quadratic in the limit because of their attention layers. Kimi Linear and DeepSeek Sparse Attention are quadratic in the implementations that actually ship, with only constant-factor speedups. The widely-cited LessWrong post on this topic is the prior that technically literate observers are starting from. SubQ's marketing is essentially a direct rebuttal of that prior. Rebuttals require evidence the public can examine.

In August 2024, Magic.dev announced a one-hundred-million-token context model called LTM-2-mini, claimed roughly thousand-times efficiency advantage, and raised more than five hundred million dollars including from Eric Schmidt and Jane Street. As of early 2026, twenty-one months later, there is no public evidence the model is used outside Magic itself. No externally usable product. No arXiv preprint. No third-party benchmarks. No reproductions.

The structural similarity to SubQ's pitch (long context, ~1,000x claim, no open weights, partnership with a GPU provider as the main public proof point, headline numbers no third party can run) is what triggered the "AI Theranos" framing on launch day. The parallel is not a verdict. It is a prior. The two timelines, side by side, make the resemblance hard to miss.

Verdicts below are not allegations. They are read priors. Each claim has a status (third-party verified, self-reported, or single-run only) and a smell-test rating against what is structurally plausible for the architecture being described. Nothing here would be load bearing if the technical paper landed on arXiv tomorrow. As of the file date, it has not.

The pattern is consistent. Production-mode, third-party-verified numbers are believable and roughly in line with what you would expect from a long-context specialist. Research-mode, self-reported, single-run numbers are exactly the cluster the field has learned to treat as marketing until a public reproduction lands. Both reads can be true at the same time. They are.