Launching today
SubQ Snap
Long-context decision model for agent workflows
1 follower
Long-context decision model for agent workflows
1 follower
SubQ Snap is a long-context decision model for AI agents. Instead of writing text, it picks from a set of allowed answers, like which file to read, which tool to call next, or whether a test passed. It takes up to 384K tokens of text and images in one call and handles 100K-token requests in under a second. Costs $0.05 per million input tokens, with output free. In Public Preview, capacity is limited.
Free Options
Launch tags:Artificial Intelligence
Launch Team
ClarifyThe full-stack AI CRM for modern GTM teams
Promoted
SubQ Snap is a long-context decision model for agentic workflows from Subquadratic (SubQ). Instead of writing text, it picks from a set of allowed answers, and it’s in Public Preview.
Problem: Agents make small decisions all the time, such as which file to read, which tool to call next, or whether a test passed. Using a full text-generating model for each one costs time and money, because the model produces text the application never needs.
Solution: Snap takes a question and a fixed list of allowed answers, and gives each answer a probability. It doesn’t generate text at all. SubQ says that removes the latency and cost of producing text the application doesn’t need.
What makes it different: the long context. Per SubQ, it reads up to 384K tokens of text and images in one call, with no truncation or chunking, and handles requests up to 100K tokens in under a second. That lets it decide over a long codebase, a document or a history in one pass. SubQ says its SSA (Subquadratic Sparse Attention) reduces attention compute cost from quadratic to linear complexity.
Key features:
Up to 384K tokens of text and images in one call
Requests up to 100K tokens in under a second
Multimodal: screenshots, charts and scanned pages alongside text
Built for routing, tool picking, checking and stopping in agent loops
$0.05 per million input tokens, with output tokens free
100M free tokens for the first 10,000 users, per the SubQ homepage
Benchmarks: these are SubQ’s own results, not independent ones. It reports 99.8 on Decision RULER at 32K tokens, and an enterprise average of 67.6 versus 61.6 for the comparison model across 16 benchmarks. Snap leads on 13 of the 16. The post doesn’t say what the comparison model is, so treat the comparison with care. See the full benchmark charts.
Who it’s for: developers building agents that make many routing or checking decisions over long inputs, such as support triage, code context curation or recommendations.
Try it in the SubQ console · Read the docs · Read the launch post