A twelve-million-token context claims a thousandth the compute

Claim

Subquadratic announced a twelve-million-token context model demanding nearly a thousand times less compute, with its sparse attention reaching 65.9% on a long-context benchmark against Opus 4.6's 78%, while Google's multi-token drafters delivered threefold speedups with no quality loss.

Reported metric: 12M tokens / ~1000x

Reporting date: 2026-05-06 (not necessarily the event date)

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

• Welcome to May 6, 2026 (theinnermostloop.substack.com)