FULL STORY

SubQ Snap: From Launch to Top of Decision Model Index

Subquadratic launched SubQ Snap on Oct 13 as a long-context decision model for enterprise agent workflows. Ten days later, SubQ Snap 1.1 topped the decision model index with 384K context and low hallucination.

2026-10-13 ~ 2026-10-23 · 2 episodes · 26 posts

Episode 1 · Subquadratic launches SubQ Snap, a 384K-context decision model with flat per-token compute (2026-10-13, 24 posts)

On October 13, startup Subquadratic launched SubQ Snap, a multimodal long-context 'decision model' positioned as the fast-decision step in enterprise agent workflows—choosing tools, opening files, checking results, deciding UI form—leaving hard problems to general LLMs like GPT. Its headline claim: per-token compute stays constant as context grows. All 24 posts in this cluster are reposts with consistent parameters and no independent verification.

Confirmed

  • Product: SubQ Snap by Subquadratic, for enterprise agent workflows; some reposts mention an SSA (sub-quadratic architecture) basis
  • 384K-token multimodal context, claimed 12x larger than competitor Jev, roughly three novels of text plus images; supports reading inputs at the 256K-token level
  • Constant per-token compute as context scales; reposts cite sub-second decisions at 100K tokens, and a separate '2-3 seconds per decision' figure
  • Pricing: $0.05 per million input tokens, output free; first 10,000 users get 100M tokens free
  • Decision Index v0.2.1 score of 61.7 vs Jev's 57.9 across 16 tests
  • Reads full conversations each turn (messages, PDFs, images, entire GitHub repos) and can directly decide UI presentation
  • Official demo played the Chrome dinosaur game; demo case located a root cause in a 3.4 MB training crash log in 0.55 seconds at $0.0002 cost (compared against Claude); a repost credits developer kevisyoung citing the official account: 'scheduled all of LA Tech Week for 1 cent'

Unconfirmed

  • The 'constant per-token compute' claim and all benchmark/demo figures are official claims without independent verification

Why it matters

If the constant per-token compute claim holds, SubQ Snap could sharply cut inference costs for long-context workloads, making large-scale deployment of agents over whole code repositories and long documents far more economical; free output and free credits for early users lower the barrier further. Its 'decision model' positioning targets routing and UI decisions in agent pipelines rather than general generation. Since this cluster consists entirely of reposts, technical details and benchmark numbers should be checked against the original official announcement.

4 more related posts →

Episode 2 · SubQ Snap 1.1 Tops Decision Model Index with 384K Context (2026-10-23, 2 posts)

Subquadratic released SubQ Snap 1.1, an enterprise decision model with 384K context, constant compute per token and low hallucination rates, topping the Artificial Analysis Decision Index v1.0 with a score of 65.