Subquadratic Snap targets long-context decisions with 384k token window

San Francisco-based startup Subquadratic has launched SubQ Snap, a new decision model designed to process up to 384,000 tokens in a single request. The release marks a significant expansion in context capacity compared to recent competitors, aiming to handle complex reasoning tasks across large codebases and document sets without chunking.

What happened

The launch follows a surge of interest in decision models, a specialized class of AI systems that select from predefined answers rather than generating open-ended text. TypeSafe’s Jev model, released in mid-September with a 32,000-token limit, sparked this trend, prompting major tech firms including OpenAI, Microsoft, and Amazon to enter the space. Subquadratic positions Snap as a direct evolution of this concept, claiming it can process twelve times the context of Jev.

Subquadratic CEO Alex Whedon states that while Jev inspired Snap, the new model is built to take substantial work away from expensive frontier large language models (LLMs). Instead of using massive general-purpose models for every step in an AI agent’s workflow, Snap handles the administrative heavy lifting. This includes selecting tools, identifying relevant files, or evaluating test results. The company argues that current decision models struggle when necessary information exceeds their small context windows, forcing developers to discard potentially critical data.

Snap is currently available in public preview. The company prices the service at $0.05 per million input tokens, with no charge for outputs. Whedon explains that because the model only reads input and returns probability scores for allowed answers, the computational cost is heavily skewed toward processing the context rather than generating text. A full 384,000-token request costs approximately two cents.

How it works

Snap relies on Subquadratic Sparse Attention (SSA), an architecture the company introduced in May. Traditional transformer models compute attention quadratically, meaning the computational load increases dramatically as the input length grows. SSA changes this relationship so that attention requirements grow linearly with context length. This allows the model to process hundreds of thousands of tokens efficiently without the exponential cost spike seen in standard architectures.

The model is designed to eliminate fragile preprocessing steps like aggressive summarization or fine-grained chunking. When an AI agent needs to fix a bug, it can feed entire modules, logs, and documentation into Snap at once. The model analyzes the full context to make a decision, such as which file to edit or which tool to call next. While retrieval systems are still needed for corpora larger than 384,000 tokens, Whedon notes that the larger window allows for coarser retrieval, pulling in whole documents instead of carefully trimmed snippets.

Key details

• Snap supports a context window of up to 384,000 tokens for text and images, though the company has tested it at one million tokens internally.
• The model processes 100,000 tokens and returns a decision in approximately one second.
• On the Decision LCR benchmark, which tests reasoning across multiple documents, Snap scored 74.0 compared to Jev’s 48.0.
• Subquadratic reports that Snap outperformed Jev on 13 out of 16 enterprise-oriented benchmarks, with an average score of 67.6 versus 61.6.
• The public preview limit of 384,000 tokens is due to current computing capacity; supporting one million tokens publicly would require 2.6 times more compute.
• Snap is not yet listed on the latest Decision Index leaderboard because the current version focuses on open-weight models and includes private tests.

Why it matters

For engineers building AI agents, the primary benefit is reduced complexity and cost. Currently, managing context involves intricate pipelines of retrieval-augmented generation (RAG), summarization, and chunking. These steps often introduce errors by omitting subtle but important details. By allowing agents to pass large swaths of raw data directly to a specialized decision model, developers can simplify their architectures. This reduces the risk of losing critical context during preprocessing.

Cost efficiency is another major factor. Frontier LLMs are expensive to run for every minor decision in an agent’s execution loop. Offloading these frequent, low-level choices to a cheaper, faster model like Snap can significantly lower operational expenses. If Snap can maintain high accuracy while processing large contexts, it enables agents to operate more autonomously over larger datasets without hitting token limits or budget caps.

However, the claims rest largely on internal benchmarks. While Subquadratic has released code for its adapted long-context evaluations, independent verification is still pending. Developers should view the performance advantages as promising but unproven in third-party settings until more external data becomes available.

What you can do

• Evaluate whether your current agent workflows involve frequent context truncation or complex chunking logic that could be simplified by a larger window.
• Test Snap’s public preview on specific tasks like codebase navigation or multi-document classification to compare latency and accuracy against your current stack.
• Calculate the cost difference between using frontier LLMs for decision steps versus paying $0.05 per million input tokens for Snap.
• Review the released evaluation code to understand how Subquadratic measures long-context retention and decide if those metrics align with your use cases.
• Monitor the upcoming updates to the Decision Index leaderboard to see if Snap appears in future independent rankings.
• Consider hybrid approaches where Snap handles broad context analysis while smaller, faster models handle immediate, local decisions.