A large language model released on May 5, 2026 by Subquadratic Inc. (US). It uses a proprietary attention mechanism called Subquadratic Sparse Attention (SSA), which curbs the compute that normally explodes as input grows, so that an entire repository or years of documents can be reasoned over in a single prompt. Where a standard Transformer computes relationships between every pair of words, SSA selects only the most relevant pairs and computes those. As of August 2026 it is not generally available; access is granted through an early-access request form on the official site.
Key Features
- Context in the millions of tokens: The company cites up to 12 million tokens in its research model and roughly 1 million tokens in the API currently offered. Demonstrations include processing the entire Python standard library (about 5.1M tokens) and several years of financial filings (about 2.9M tokens) in one pass, without chunking
- Reduced compute via SSA: Instead of dense attention that compares every token with every other token, each token attends only to the tokens most relevant to it. The company reports 7.8 PFLOP to process 2M tokens, roughly 1/128th of the 1,008 PFLOP it attributes to frontier models
- Throughput on long inputs: A published figure puts it at 52x faster than FlashAttention at 1M tokens. By design, the gap against existing approaches widens as inputs get longer
- Long-context accuracy benchmarks: 99.12% on multi-fact retrieval in RULER 128K, and 100% on single-fact retrieval (needle-in-a-haystack) at 1–2M tokens. It also reports 85.4% on GPQA Diamond and 89.7% on LiveCodeBench (all figures published by the company)
- OpenAI-compatible API and developer tools: The API is offered in an OpenAI-compatible format, and a CLI agent called SubQ Code plus a retrieval-oriented SubQ Search are listed as part of early access
- Ongoing model updates: A lighter variant, SubQ 1.1 Small, was released in June 2026, showing continued updates to the model line after launch
Pricing
| Plan | Price | What’s included |
|---|---|---|
| Early access (API) | Not disclosed / contact required | Apply via the form on the official site; access to the API, SubQ Code, and SubQ Search after review |
| Enterprise | Contact required | Aimed at organizations processing large repositories, contract archives, and financial documents |
Pricing is as of August 2026. No public rate card (per-token pricing) is published on the official site; pricing is communicated individually after you apply. Check the official site for the latest information.
Note that the company and press coverage cite cost comparisons such as “running a 128K long-context evaluation for about $8.” That is an estimate for a benchmark run, not a published usage-based rate.
Pros & Cons
✅ Pros
- Handling millions of tokens without splitting them can remove the need to build out chunking and retrieval pipelines like RAG
- The compute advantage grows with length, making bulk analysis of large codebases and long document archives easier to keep affordable
- The OpenAI-compatible API keeps migration cost low for existing client code
- Published long-context retrieval benchmarks are high, and the design directly targets the “loaded but not actually read” problem
⚠️ Cons
- Not generally available as of August 2026; an early-access application and review are required
- No published per-token price, making it hard to estimate cost before adoption
- The headline 12 million tokens applies to the research model and differs from the API’s limit
- It is specialized for long-context processing rather than general chat; there is little information on image generation or everyday conversational use
- As a young vendor, its long-term availability and update cadence are unproven
Comparison with Similar Services
| Criteria | SubQ | Gemini | Claude |
|---|---|---|---|
| Provider | Subquadratic Inc. | Anthropic | |
| Focus | Efficient ultra-long-context processing | General multimodal + long context | Natural writing and long-context processing |
| Stated context | 12M in research model / ~1M in API | Around 1M tokens | Around 1M tokens |
| Architectural trait | Sparse attention (SSA) to cut compute | Primarily dense attention | Primarily dense attention |
| Availability | Early access by application | Generally available | Generally available |
| Public pricing | Not disclosed | Published | Published |
Who Is It For
- Development teams that want an entire monorepo or large codebase analyzed in one pass, without splitting it
- Business users who need cross-cutting analysis of long document sets such as contract archives or annual financial reports
- Engineers who would rather put everything in the context window than spend effort tuning RAG retrieval accuracy
- Organizations where the inference cost of long-context processing has become an operational bottleneck
Summary
SubQ aims squarely at one question: how to read long inputs cheaply and quickly. SSA slows the growth of compute relative to input length, enabling a workflow where whole repositories or years of material are handled without chunking. That said, as of August 2026 it remains in early access with undisclosed pricing, so any adoption decision requires confirming the individual terms offered after you apply. It is worth considering if long-context processing has hit a cost ceiling for you.