Gemini 4 Argon starts with cyber defenders, not a public API

Google announced Gemini 4 Argon on September 30, but most developers cannot call it yet. The initial rollout goes to a set of trusted cyber defenders through Google DeepMind’s Fairwind Program. Google says paid API customers and Google AI Ultra subscribers will be among the first in a broader release, without giving a date. That makes this a model announcement and a controlled deployment, not a general API launch.

The standout specification needs careful reading: Google says Argon’s output token limit rises from 64K to one million tokens. This is not an announcement of a one-million-token input context window. The larger output budget is meant to let the model continue a long reasoning and generation trajectory, but the announcement does not establish what a typical user will receive in a single visible response, or publish an Argon API model ID.

The work Google wants Argon to do

Google pitches Argon for sustained software engineering, enterprise research and defensive security. Its launch post reports 77.9% on DeepSWE v1.1, a long-horizon software-engineering evaluation, and 51.3% on AutomationBench, which tests end-to-end business tasks. For vulnerability remediation, it says Argon ties for first at 68% on CWE-bench v1. Google also reports 91.7% on LVBench for long-video understanding and calls Argon the leading model on the Vals Index, which combines finance, coding, legal and tax work. These are Google-reported results on different evaluations, not one comparable score or a guarantee that an agent will complete an unfamiliar migration or find a production vulnerability unattended.

The launch material does not provide a public Argon API model ID or a reproducible developer setup for those tests. Google’s scorecard and its broad “leading” language are vendor claims; the numbers above should be read benchmark by benchmark rather than as an independently established ranking across all frontier models.

Google offers a more concrete account of its internal use: Argon agents worked on C/C++-to-Rust migrations, including code in the Fuchsia Zircon kernel. The company says those large rewrites are still subject to automated and manual auditing, emulation tests and review before production rollout. That qualification is the useful part of the example. A model that writes a large patch does not thereby make the patch safe to ship.

The cyber deployment is unusually explicit about who gets the more capable configuration. Google says trusted defenders and its internal teams will receive Argon without cyber guardrails for defensive work; its wider release is being held while it strengthens safeguards against misuse, prompt injection and misalignment. The Fairwind page says only a set of program partners get Argon access, either as a standalone model or alongside CodeMender. Organizations must restrict access to their internal cyber, incident-response or penetration-testing teams, track users, and cannot redistribute access. Applying to Fairwind is not the same as obtaining the model.

A price announced ahead of availability

Google’s published launch terms put Argon’s future introductory rate at $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off the input rate. The footnote says the rates become $4 input and $20 output per million after the introductory period, but does not give that period an end date. Those numbers should not be mistaken for a tariff available to every developer today: Argon was not listed on Google’s public Gemini API pricing or model directory pages at this draft’s September 30 check.

For teams deciding whether to adopt it, the next meaningful test is access to an actual model ID, documented limits and repeatable evaluations on their own workflows. Until then, the announcement establishes an ambitious output ceiling, vendor-reported results and a narrow defensive rollout. It does not establish that a public developer can run Argon now.