The grenade thrown on 5 May 2026
On 5 May 2026, a small startup based in Miami called Subquadratic came out of its "stealth" phase with a claim that landed like a grenade in an industry accustomed to incremental progress. The company claims to have built the first commercial language model based entirely on a sub-quadratic architecture, meaning its compute requirements grow linearly (not quadratically) with context length.
To grasp the scale of this claim, you need to understand the technical background. Since the publication of the famous "Attention is All You Need" paper in 2017 by Google researchers, every major AI model has rested on the same architecture: the transformer. OpenAI's GPT-5.5, Anthropic's Claude Opus 4.7, Google's Gemini 3.1 Pro, Meta's Llama, Mistral in France: all transformers. If SubQ really works, it would be the biggest architectural break since 9 years.
This article unpicks what is going on, what we actually know, and why the AI community is deeply divided on the implications. Following on from our comparison of the major 2026 models, we look here at a player that could redefine the frontier over the next 18 months.
The fundamental problem SubQ claims to solve
Transformer attention operates with what is known as "quadratic" complexity, written O(n²) in mathematical notation. In practice, this means that doubling the length of an input text multiplies compute costs by four. Tripling the length multiplies them by nine. This exponential explosion is why "1M token context" models are very expensive for long contexts and often degrade in quality as you approach their limit.
Several teams have tried to work around this problem over the years with different approaches:
- Mamba and state space models: alternative architectures that scale linearly, but often lose accuracy when retrieving distant information.
- Sparse attention: only computes attention over certain token pairs deemed important. An approximation that can miss crucial connections.
- Sliding window attention: limits attention to a fixed window, but prevents long-distance connections.
- Hybrids: combinations of the above techniques, which generally retain a quadratic component.
Until now, none had managed to maintain "frontier" quality (equivalent to the best transformer models) while breaking the quadratic barrier. That is precisely what SubQ claims to have solved.
The architecture is called SSA (Subquadratic Sparse Attention). The principle: a content-dependent sparse routing mechanism that computes exact attention only over the tokens that genuinely matter. Unlike approximation-based approaches, SSA does not merely estimate attention on the important tokens; it ignores it entirely on tokens deemed irrelevant. The technical result: complexity drops from O(n²) to near-linear, while preserving accuracy on the important connections.
The claimed figures
Subquadratic publishes several precise figures to back up its claims. These are the company's official figures, so they should be treated with caution pending independent validation.
| Metric | Claimed value |
|---|---|
| Maximum context | 12 million tokens |
| Reduction in attention computation at 12M tokens | 1,000 times less |
| Speedup vs FlashAttention at 1M tokens | 52.2 times faster |
| Price per million tokens (input) | $0.50 (vs $1.50-15 for competitors) |
| RULER score at 128k tokens | 95.0% |
| SWE-Bench Verified score (coding) | 81.8% |
| Inference speed | 150+ tokens/second |
RULER (the long-context reference test, which measures the ability to carry out multiple information retrievals, track variables and aggregate documents) is probably the most relevant benchmark for evaluating a long-context model. A score of 95% at 128,000 tokens is excellent. SWE-Bench Verified (a coding benchmark) at 81.8% puts SubQ on a par with Claude Opus 4.7, which we documented in our article on Claude Opus 4.7.
The economics that have the competition rattled
The prices announced by SubQ are brutal. $0.50 per million input tokens, $1.50 per million output tokens. For a quick comparison with the established players (based on our analysis in the Gemini Flash vs competitors comparison):
- Gemini 3.5 Flash: $1.50 / $9
- GPT-5.5: $5 / $30
- Claude Opus 4.7: $15 / $75
SubQ is therefore 3 times cheaper than Gemini 3.5 Flash, 10 times cheaper than GPT-5.5, and 30 times cheaper than Claude Opus 4.7. If these prices hold (and if the quality is confirmed by independent benchmarks), it is an economic earthquake for the industry. For long-context use cases (analysing entire codebases, reading complete libraries, processing large legal or medical documents), SubQ could become the default choice.
The use cases this unlocks
If we take the claims seriously, 12 million tokens of usable context changes the nature of what can be done with an LLM. A few concrete examples of what this makes technically possible:
Loading an entire codebase. The Linux source code runs to around 28 million lines. It would probably take 2-3 SubQ instances to cover it. But medium-sized projects (Kubernetes, React, Postgres) fit easily within 12M tokens. An AI agent could reason over all the code in a single pass, without needing RAG (Retrieval Augmented Generation: a technique that involves retrieving relevant chunks of a corpus to feed to the model).
Reading a complete book in context. An 800-page novel is around 200,000 tokens. A library of 50 books fits within 12M tokens. For an author, a researcher or a lawyer who has to work with massive corpora, this is revolutionary.
Analysing entire medical records. A patient's complete history since birth can fit in the context. For complex diagnoses, that is invaluable. As we explored in our article on the future of medicine, this kind of capability is crucial for personalised medicine.
Maintaining genuinely long conversations. An AI assistant could remember hundreds of hours of dialogue with its user without losing the thread. That is precisely what today's assistants sorely lack.
Reducing dependence on RAG. RAG has become the standard way of giving models access to external knowledge. But it is complex to set up, fragile, and costly in infrastructure. With 12M tokens, many RAG applications become obsolete: the relevant documents are loaded directly into the context. As we explored in our article on RAG, this technology could be largely replaced if SubQ lives up to its promises.
The criticism mounting in the community
Despite the enthusiasm, several serious criticisms emerged in the hours following the announcement. Let us list them honestly.
Criticism 1: few benchmarks published. SubQ has published results on only 3 benchmarks: RULER, SWE-Bench Verified, and MRCR v2 (another long-context test). All are focused on what the model is supposed to do well: long-context retrieval and coding. No results published on the classic benchmarks for general reasoning (MMLU, ARC-AGI), mathematics (AIME), multilingualism, or safety. The full model card is listed as "coming soon" on the company's site.
Criticism 2: the choice of comparisons. The benchmarks selected compare SubQ against transformer models on their weaknesses (long context), not on their strengths. It is the equivalent of comparing a Formula 1 car and a tractor on traction in mud: the result is known in advance.
Criticism 3: non-transformer promises have already come unstuck. Mamba (a competing architecture, presented in 2023) was also announced as "the end of the transformer". The first results were impressive, but extended evaluations revealed significant limitations, and Mamba remains a minority approach in 2026. Several researchers note that SubQ could follow the same trajectory.
Criticism 4: the team is small. Subquadratic has around twenty people, against thousands at OpenAI, Anthropic, Google. If the idea is as good as claimed, why has no major lab developed it? The team is made up of PhDs from Meta, Google, Oxford, BYU, ByteDance, Adobe, Cambridge, so the talent is there. But the capacity for long-term execution remains to be proven.
Criticism 5: 29 million dollars is little. To train and operate a frontier model, you typically need several hundred million dollars in compute. SubQ has raised 29M in seed, which is high for a stealth startup but low for a frontier player. The question: how long will it hold out financially, and who might buy it?
The real validation will come from independent evaluations. Several university teams (Stanford, MIT, Carnegie Mellon, Oxford) have already requested access. The first results are expected in June-July 2026. Artificial Analysis, the independent firm that benchmarks all public models, has scheduled its evaluation for late May. If SubQ confirms its figures, the revolution is real. If the public benchmarks significantly degrade the official results, it will be the umpteenth case of an AI company that oversold its capabilities, as we documented in our article on faked AI demos.
Why this story matters beyond SubQ
Whether or not SubQ keeps its promises, the story itself reveals several important dynamics in the AI market in 2026.
First, architectural diversification is accelerating. For 9 years, all frontier models were transformers. Today, we see SubQ with its SSA, Mamba with its state space models, several hybrids, and probably others in development. This diversity is healthy: it reduces the risk that an architectural flaw in the transformer blocks the entire sector.
Second, startups can still challenge the giants. In 2024-2025, many observers thought that AI had become a field where only OpenAI, Anthropic, Google and Meta could innovate, for lack of capital. SubQ demonstrates that small but brilliant teams can still create fundamental breakthroughs. This is rather reassuring for the health of the ecosystem, in line with our analysis in our article on the startups that outperform.
Third, unit economics can still drop drastically. If SubQ delivers 30 times cheaper than Opus 4.7 for similar tasks, it forces the entire industry to review its prices. As we analysed in our article on the AI bubble, this commodification is accelerating and could weigh on the margins of today's leaders.
Fourth, AI infrastructure is being transformed. If sub-quadratic models take hold, GPU requirements will change. Nvidia's H100/H200 chips are optimised for transformer attention computations. A different architecture could reshuffle the hardware deck. AMD, which recently scored points with Zyphra ZAYA1 trained on its chips, could benefit.
What should be done in practice?
For different profiles, here are the immediate concrete implications.
For an independent developer. If you have a long-context use case (code analysis, processing long documents), test SubQ via their early access API available on subq.ai. No need to wait for the community to validate it: try it on your own cases. The risk is low (the API is billed on usage), the potential gain is high.
For a company. Don't abandon your RAG infrastructure overnight. As we explored in our article on RAG, that technology remains relevant. But start evaluating SubQ in parallel for your long-context use cases. If the quality holds up, you could save massively on infrastructure over the next 12 months.
For an investor. Keep an eye on the independent evaluations due in June-July. If they confirm the claims, Subquadratic will probably become an acquisition target for one of the giants (Meta, Google, or Anthropic) within the next 18 months. If not, the market will quickly return to its usual valuations.
For an AI researcher. Subquadratic's technical paper is in the process of being published. When it comes out, it will probably be one of the most important reads of 2026. Beyond SubQ itself, the SSA approach could inspire other teams to explore alternative architectures.
The 18-month horizon
Looking 18 months ahead, three scenarios are possible.
Optimistic scenario: SubQ delivers on its promises. Independent benchmarks confirm the claims, several companies adopt the architecture, OpenAI/Anthropic/Google develop their own sub-quadratic variants. The industry gradually moves to hybrid or non-transformer architectures. Probability: 30-40%.
Middle scenario: SubQ works, but in a niche. The model delivers on long context but remains weaker on general reasoning. It becomes a specialist player for long-context use cases (coding, documents, research), while transformers continue to dominate the rest. Probability: 40-50%.
Pessimistic scenario: SubQ fails the independent test. External evaluations reveal significant degradation, the company loses credibility, and the idea of a sub-quadratic architecture remains a research topic for a few more years. Probability: 15-25%.
None of these scenarios is negligible. That is probably the main reason the AI community is following SubQ so closely.
To go further, you can read our comparison of frontier models, which places SubQ in the current competitive context, and our article on how LLMs fundamentally work, which explains why architecture matters so much. The next decade of AI could hinge on these seemingly esoteric technical choices. And SubQ could be the starting gun.