• Sources: announcement, HN discussion
  • Summary: Google states an introductory price of 2 dollars per million input tokens and 10 dollars per million output tokens, rising to 4 and 20 after the introductory period, with cached input at 95 percent off the input price. It states the output token limit rises from 64K to 1M, which is how much a single generation can produce rather than how much the model can read, and the page's own generated summary describes a 1 million token limit without that qualification. Google credits Argon agents with applied results it does not substantiate outside the post: 32K lines of SIMD code replaced in the libgav1 Rust port, running 2.7 times faster than that port with identical video output, C and C++ to Rust migrations of up to 800K lines for the Fuchsia Zircon kernel, and a fleet-wide memory pass that freed over 300 TiB. The benchmark figures are Google's own on the same footing, including DeepSWE v1.1 at 77.9 percent, AutomationBench at 51.3 percent, LVBench at 91.7 percent, and CWE-bench v1 at 68 percent, with no independent reproduction available on announcement day.
  • Why it matters: Nobody outside the Fairwind cohort can use the model yet, so the price and the output limit are what a reader can plan against while the migration and memory results stay unverified, and shipping a build without cyber guardrails to trusted defenders and Google's own internal teams is a deliberate asymmetry with the developers, enterprises, and consumers who get it later.
  • Follow-up: Track general availability, whether the post-introductory price holds at 4 and 20, and whether the build without cyber guardrails reaches anyone beyond the Fairwind cohort.