SubQ Snap
A decision model that does not generate text but instead computes a probability distribution over given candidate answers, designed specifically for agent routing, tool selection, verification, and stop decisions.
Timeline (last 90 days)
SubQ Snap is a decision model released by the research organization Subquadratic; it does not generate text, but instead returns a probability distribution for a given set of candidate answers.
SubQ Snap scored 67.0 in the When2Call test, losing to a competitor's 74.6.
SubQ Snap is priced at $0.05 per million input tokens, with a context window of 384K.
Subquadratic claims that SubQ Snap's Subquadratic Sparse Attention (SSA) can reduce attention cost from quadratic to linear, returning 100K tokens within 1 second, but these figures have no hardware specification or third-party verification.
SubQ Snap officially claims to lead on 13 of 16 enterprise tasks, with phishing detection improving from 25.1 to 47.1 and code output prediction improving from 57.1 to 73.8.
SubQ Snap's official headline result is a Decision RULER score of 96.3 running in a 1M token context, but the product itself only supports a 384K context.
All of SubQ Snap's performance data comes from Subquadratic itself and has not been independently tested by an external lab, and its long-context tests are its own modified versions of RULER and LCR.
Searching 'subq snap' on Reddit cannot find the model; the results are mainly a forum on subcutaneous fluids for cat kidney failure, reviews of SNAP-8 peptide anti-wrinkle serum, and discussions of spring syringes.
SubQ Snap's model ID is 'subq-snap-preview', indicating it is in public preview and should not be relied on as a stable version.
All Facts (9)
SubQ Snap is a decision model released by the research organization Subquadratic; it does not generate text, but instead returns a probability distribution for a given set of candidate answers.
50%UnverifiedSubQ Snap scored 67.0 in the When2Call test, losing to a competitor's 74.6.
50%UnverifiedSubQ Snap is priced at $0.05 per million input tokens, with a context window of 384K.
50%UnverifiedSubquadratic claims that SubQ Snap's Subquadratic Sparse Attention (SSA) can reduce attention cost from quadratic to linear, returning 100K tokens within 1 second, but these figures have no hardware specification or third-party verification.
50%UnverifiedSubQ Snap officially claims to lead on 13 of 16 enterprise tasks, with phishing detection improving from 25.1 to 47.1 and code output prediction improving from 57.1 to 73.8.
50%UnverifiedSubQ Snap's official headline result is a Decision RULER score of 96.3 running in a 1M token context, but the product itself only supports a 384K context.
50%UnverifiedAll of SubQ Snap's performance data comes from Subquadratic itself and has not been independently tested by an external lab, and its long-context tests are its own modified versions of RULER and LCR.
50%UnverifiedSearching 'subq snap' on Reddit cannot find the model; the results are mainly a forum on subcutaneous fluids for cat kidney failure, reviews of SNAP-8 peptide anti-wrinkle serum, and discussions of spring syringes.
50%UnverifiedSubQ Snap's model ID is 'subq-snap-preview', indicating it is in public preview and should not be relied on as a stable version.
50%