SubQ by Subquadratic: 12M Context LLM Reviewed (2026)
By Loïc Jané · Founder, Fleece AI
SubQ by Subquadratic: The First 12M-Token Subquadratic LLM Reviewed (2026)
At a Glance: SubQ is a new large language model from Miami startup Subquadratic, launched May 5, 2026 with a $29M seed round. According to SiliconANGLE's launch coverage, SubQ uses a novel Subquadratic Selective Attention (SSA) architecture and claims a 12-million-token context window — roughly 9 million words or 120 books — while running 52× faster than dense attention at 1M tokens. The catch: as VentureBeat reports, benchmarks are self-reported, no public technical paper exists, and researchers are demanding independent proof. This review explains what's real, what's unverified, and what it means for AI agent platforms like Fleece AI. Published May 6, 2026; updated September 28, 2026.
Update (September 2026): on June 16, 2026, Subquadratic released SubQ 1.1 Small with a technical report and model card, now describing SSA as "Subquadratic Sparse Attention." The report claims 98% single-fact retrieval at 12M tokens and 64.5x less attention compute than dense attention at 1M tokens. Subquadratic also published a benchmark evaluation by Appen, which Subquadratic commissioned. The model was built by replacing dense attention in an existing open-weight model with SSA. As of September 2026, API access is still by early-access request, pricing is not public, and no lab unaffiliated with Subquadratic has published a reproduction. SubQ is not available on Fleece AI.
Table of Contents
- What Is SubQ?
- What Is Subquadratic Selective Attention (SSA)?
- Performance Claims (Self-Reported)
- Why the Research Community Is Skeptical
- SubQ vs Other Frontier Models
- What 12M Tokens Means for AI Agents
- How Long-Context Models Fit Into Fleece AI
- Will SubQ Land on Fleece AI?
- FAQ
Key Takeaways
- SubQ is the first publicly launched LLM built on a fully subquadratic attention architecture — the company calls it Subquadratic Selective Attention (SSA) — meaning compute scales linearly, not quadratically, with context length.
- The launch model SubQ 1M-Preview ships a 12-million-token context window (50M planned), which Subquadratic claims is roughly 9 million words or ~120 books in a single context.
- Performance claims include 52× faster inference than dense attention at 1M tokens, 92.1% needle-in-haystack at 12M tokens, and 83 on MRCR v2 — but as VentureBeat notes, every benchmark is self-reported with no third-party verification or public technical paper.
- Subquadratic raised $29M seed from investors including Justin Mateen (Tinder co-founder) and early backers of Anthropic, OpenAI, Stripe, and Brex; products in private beta include the API, SubQ Code (coding agent), and SubQ Search.
- Fleece AI tracks frontier model launches but evaluates new models on real-world agent reliability before adding them to plan tiers — SubQ is not on Fleece AI as of May 2026 and won't be until independent benchmarks confirm the architectural claims.
What Is SubQ?
SubQ is the first frontier large language model from Subquadratic, a Miami-based AI startup that emerged from stealth on May 5, 2026 with a $29M seed round. The launch model — SubQ 1M-Preview — claims to be the first commercial LLM built on a fully subquadratic attention architecture, meaning compute requirements grow linearly with context length instead of quadratically.
According to Subquadratic's launch announcement, the model handles a 12-million-token context window with 92.1% accuracy on needle-in-a-haystack retrieval, runs at 52× the speed of dense attention at 1M tokens, and reduces compute requirements by nearly 1,000× at the full 12M context length. The company has positioned itself as challenging "the mathematical constraint that has defined every major AI system since 2017" — the quadratic cost of self-attention in transformer architectures.
What Is Subquadratic Selective Attention (SSA)?
Subquadratic Selective Attention is the company's proprietary attention architecture. According to Subquadratic's technical post on SSA, SSA scales linearly in both compute and memory with respect to context length — sidestepping the O(n²) cost that makes a 12M-token window prohibitively expensive on standard transformers.
The promise of subquadratic attention is not new — Mamba, RWKV, Hyena, and other linear-attention architectures have explored variants since 2023. What's distinctive about Subquadratic's claim is the combination of (a) full subquadratic scaling, (b) competitive accuracy on long-context retrieval, and (c) frontier-class benchmark performance. If the architecture delivers as advertised, it would be the first time a sub-quadratic model competes with dense transformers across the board — not just in latency-critical niches.
The skepticism in the research community is not about whether subquadratic attention can work — it's about whether this particular implementation matches what's claimed. There is currently no public paper describing SSA in enough detail to reproduce.
Performance Claims (Self-Reported)
Subquadratic has published the following self-reported benchmarks alongside the SubQ launch:
| Benchmark | SubQ Claim | Context |
|---|---|---|
| Speed at 1M tokens | 52× faster than dense attention | Self-reported, no third-party run |
| Compute reduction at 12M | ~1,000× lower than frontier models | Self-reported |
| Needle-in-haystack at 12M | 92.1% accuracy | Self-reported |
| MRCR v2 | 83 (claimed +9 points vs OpenAI) | Self-reported |
| Context window | 12M tokens (50M roadmap) | Reported by The New Stack |
| Funding | $29M seed | Confirmed by SiliconANGLE |
| Best For | Long-document QA, codebase understanding, persistent agent memory (if claims hold) | Marketing positioning |
| Pricing | Not disclosed at launch | Private beta only |
Until independent labs run SubQ on standardized benchmarks, treat every performance number as a Subquadratic claim, not a verified fact.
Need a working AI agent platform today? Fleece AI runs Claude Sonnet 5.5 (the default), Claude Haiku 5.5, Mistral Medium 3.5 and, on Business, Claude Opus 5.5 — all production-tested. Start at fleeceai.app.
Why the Research Community Is Skeptical
Three reasons skepticism was louder than usual for SubQ at launch:
- No public technical paper. Frontier model launches typically ship a paper or model card with enough detail for outside researchers to reason about. SubQ launched with a marketing post and benchmark numbers — that is not how the field validates claims. (A technical report for SubQ 1.1 Small followed on June 16, 2026.)
- The Magic.dev precedent. As VentureBeat reminds readers, Magic.dev announced a 100M-token context model (LTM-2-mini) with similar 1,000× efficiency claims in August 2024 and raised ~$500M on the strength of those claims; as of early 2026, there is no public evidence of LTM-2-mini being used outside Magic.
- Self-reported benchmarks beating frontier labs. SubQ's claim of beating OpenAI by 9 points on MRCR v2 is the kind of result that, if real, would dominate AI conferences. The default Bayesian prior on extraordinary claims is "wait for replication."
None of this means SubQ isn't real. It means responsible coverage — and responsible model selection for production agents — requires waiting for verification.
SubQ vs Other Frontier Models
Comparison as of May 2026, with context windows, availability and prices corrected in September 2026.
| Dimension | SubQ 1M-Preview | GPT-5.2 | Claude Opus 4.7 | Gemini 3.1 Pro |
|---|---|---|---|---|
| Best For | Long-document workflows (if validated) | Agent tool calling | Reasoning + coding | Multimodal long-context |
| Released | May 5, 2026 (private beta) | December 11, 2025 | April 16, 2026 | February 2026 |
| Context window | 12M tokens (claimed) | 400K tokens | 1M tokens | 1M tokens |
| Architecture | SSA (subquadratic) | Transformer | Transformer | Transformer |
| Independent verification | Vendor-commissioned only (Appen, June 2026) | Standard | Standard | Standard |
| Tool calling reliability | Unknown | Strong | Strong | Improving |
| Fleece AI availability | Not available | All plans | Business tier | Not on Fleece AI |
| Pricing transparency | Not disclosed | Public | Public | Public |
| Pricing (per 1M tokens) | TBA | $1.75/$14 | $5/$25 | $2/$12 (≤200K context) |
For more on each model, see our reviews of GPT-5.2, Claude Opus 4.7, and Gemini 3.1 Pro.
What 12M Tokens Means for AI Agents
If SubQ's claims hold, a 12M-token context window genuinely changes how AI agents are built. Today, agent platforms work around small windows with Retrieval-Augmented Generation (RAG): documents get chunked, embedded, indexed, and the agent retrieves the top-k chunks at inference. RAG works, but it adds latency, infrastructure (vector DBs, embedders, re-rankers), and brittleness — agents miss context that isn't surfaced by the retriever.
A genuine 12M-token native context could:
- Drop entire knowledge bases into the prompt at run time, eliminating retrieval brittleness for many use cases.
- Hold a full quarter of CRM activity for an account in a single agent call — letting a sales agent reason over the actual history rather than a summary.
- Persist agent memory inline across a long-running session — your manager agent can hold a full project's context without offloading to external state.
- Replace some multi-agent hierarchies with a single long-context agent — when the tool list is small but the context is huge.
If they don't hold — or if SubQ scales subquadratically only on niche tasks — the field reverts to RAG + transformer baselines and the long-context narrative ages a year.
How Long-Context Models Fit Into Fleece AI
Fleece AI agents already use long-context models in production. Claude Opus 4.7 and Opus 5.5 ship with a 1M-token window on the Business tier, and DeepSeek V4.1 Flash · EU brings a 1M-token window to every plan. For most workflows — daily digests, multi-app automation, scheduled reporting — 200K to 1M tokens is more than enough.
The use cases where 12M+ would be transformative inside Fleece AI:
- Codebase-aware coding agents — drop an entire repo in context and reason holistically. Pairs with GitHub automation.
- Cross-document research — analyze ~120 books or thousands of PDFs in one agent call.
- Persistent agent memory — replace knowledge files + conversation history with native context.
- Cross-quarter sales analysis — hold every customer interaction across multiple years inline.
These are real wins — if the model is reliable enough for production. Fleece AI does not add models to plan tiers based on benchmarks alone. Reliability on tool calling, latency on real agent runs, and stability of structured output all matter.
Will SubQ Land on Fleece AI?
Not yet. Fleece AI's model selection criteria require: (a) public availability beyond invite-only beta, (b) independent benchmark verification, (c) reliable tool calling on real agent runs, and (d) a pricing model that maps to our credit system. As of September 2026, SubQ still fails (a), (b) and (d): access is by early-access request, the only third-party evaluation was commissioned by Subquadratic, and pricing is not public.
In May we expected independent reproductions in Q3 2026 and a public API with pricing in Q4 2026. The first did not happen in Q3: Subquadratic shipped a technical report and a vendor-commissioned evaluation instead. We will reassess when a public API and unaffiliated benchmarks exist.
Until then, Claude Sonnet 5.5 (the default), Claude Haiku 5.5, Mistral Medium 3.5, and Claude Opus 5.5 and 4.7 (Business) are the production-tested options for autonomous AI agents on Fleece AI.
Frequently Asked Questions
Is SubQ available to use today?
Not publicly. According to SiliconANGLE's launch coverage, Subquadratic launched three products into private beta in May 2026 — the API, SubQ Code, and SubQ Search. As of September 2026, API access is still granted through an early-access request, SubQ 1.1 Small (June 2026) is being deployed with design partners, and there is no public pricing or self-serve access.
How does SubQ compare to GPT-5.2 or Claude Opus 4.7?
In context window, SubQ's claimed 12M tokens dramatically exceeds GPT-5.2 (400K) and Opus 4.7 (1M). On reasoning, tool calling, and production reliability, GPT-5.2 and Opus 4.7 have months of public use; SubQ has no public production track record. See the best AI models for workflow automation for the production options.
What is Subquadratic Selective Attention (SSA)?
SSA is Subquadratic's proprietary attention architecture that the company says scales linearly in compute and memory with context length, instead of quadratically. Subquadratic now calls it Subquadratic Sparse Attention. A technical report for SubQ 1.1 Small was published on June 16, 2026, but no peer-reviewed paper or independent reproduction exists as of September 2026.
Should I switch to SubQ for my AI agents?
Not yet. Pre-public-API, pre-independent-verification models are appropriate for research; production agent platforms — including Fleece AI — wait for replication and stability data. The June 2026 Appen evaluation was commissioned by Subquadratic, so it does not replace an unaffiliated reproduction. Track the situation; don't move workloads until at least one independent lab publishes benchmarks.
Why is the AI research community skeptical of SubQ?
Three reasons: (1) no public technical paper at launch (a technical report followed in June 2026), (2) the precedent of Magic.dev's similar 100M-token claims in 2024 that have not been independently verified, and (3) self-reported benchmarks that beat frontier labs by margins large enough to merit caution. As VentureBeat reports, researchers are openly demanding independent proof.
The Bottom Line
SubQ may turn out to be the most important AI model release of 2026 — or it may turn out to be the most ambitious vaporware since Magic.dev. There was no way to know in May 2026, and as of September 2026 the architecture has still not been verified by an unaffiliated lab and the model is still not publicly accessible. Production agent platforms — Fleece AI included — will watch closely and add SubQ when the evidence catches up to the claims. For now, the production-tested model lineup is what runs your real workflows.
Related Articles
- Best AI Models for Workflow Automation 2026 — the production lineup
- Long Context AI Agents Guide — context-window deep dive
- Claude Opus 4.7 on Fleece AI — Anthropic's April 2026 flagship
- GPT-5.2 review — OpenAI's former flagship, no longer offered on Fleece AI
- Gemini 3.1 Pro Review — multimodal long-context
- AI Agent Benchmarks 2026 Explained — how to read claims
- Best Autonomous AI Agents 2026 — platform roundup
- What Is Fleece AI? — platform overview
Build with production-tested models on Fleece AI — Claude Sonnet 5.5, Mistral Medium 3.5, and Claude Opus 5.5 (Business) ready today, no invite list.