SubQ Snap is a multimodal model for evaluating supplied context and returning structured decisions, such as a choice, score, or label, for agent workflows. SubQ says it supports 384K tokens and uses Subquadratic Sparse Attention so compute per token stays constant as context grows; the 384K limit and scaling claim are company claims, not independently verified here (SubQ, launch post).

SubQ Snap is a multimodal model for evaluating supplied context and returning structured decisions, such as a choice, score, or label, for agent workflows. SubQ says it supports 384K tokens and uses Subquadratic Sparse Attention so compute per token stays constant as context grows; the 384K limit and scaling claim are company claims, not independently verified here (SubQ, launch post).

The 100M free input tokens for the first 10,000 sign-ups is also stated by SubQ on its site and in the launch post. The post says output is free, paid input is $0.05 per 1M tokens, and latency varies; its reported times range from 80 ms with minimal context to 2.5 seconds at 384K (launch post).

In short, the forwarded summary matches SubQ’s announced terms, but the capacity-based sign-up limit and technical claims are not independently confirmed by these sources.