Google DeepMind releases Gemini 4 Argon in limited cybersecurity preview, with 1M-token output via Long Decode Continuation

Google DeepMind announced Gemini 4 Argon, its first model above Flash tier since Gemini 3.1 Pro in February. Access is limited to government users and trusted cyber defenders in the Fairwind Program, with no date for wider release. A new API feature, Long Decode Continuation, pauses and resumes long responses across calls, raising the output ceiling from 64K to 1M tokens. List pricing is $4 input and $20 output per 1M tokens, with a 50% introductory discount ($2/$10) and no announced end date, plus a 95% discount on cached input.

For most developers and enterprises this is a preview, not a product they can buy, so procurement or architecture decisions should wait. The practical change is the long-output feature and the pricing, which would matter for large code generation and document-scale work if they hold up. It also shows Google trying to compete again at the top end against OpenAI and Anthropic, but that is only testable once outsiders can run it widely. The cyber-defence restriction suggests Google sees real misuse risk, which may shape how future frontier releases are gated.

The "13 of 19 benchmarks" and "back at the frontier" claims are Google's own framing, and independent tests show mixed results: heavier output-token use per task and lower accuracy on some tests. The "1M-token output" headline is also overstated, since Vals measured a 262K output limit. We also only have this via an aggregator newsletter with no primary source links extracted, so treat details as unverified.

1. 01Date and terms for general availability beyond the Fairwind Program, and whether the introductory 50% discount gets an end date
2. 02Independent re-tests of the output limit (Vals saw 262K versus the claimed 1M) and cost per completed task, given the higher output-token use
3. 03Third-party benchmark results from Artificial Analysis and others against GPT-6 Astra and Claude Opus 5.5 once public access opens, including any rival price or feature responses