Gemini 4 Argon Ships With 1M Output Tokens via Fairwind

Gemini 4 Argon is Google DeepMind’s new frontier model, rolling out first to trusted cyber defenders through the Fairwind Program. The company is expanding the output token limit to an industry-leading 1M tokens, up from 64K, and pricing the model at an introductory $2 / $10 per 1M input/output tokens.

Output headroom and pricing

Per the Google blog, Argon’s 1M output token limit is meant for long-horizon workflows that need hundreds of thousands of tokens in a single trajectory. Cached input tokens are priced at 95% off the input rate. After the introductory period, list price becomes $4 / $20 per 1M input/output tokens.

Fairwind-first, phased release

Google says a phased approach is required at this capability level. Argon is going first to Fairwind partners — governments, critical infrastructure operators, and other trusted defenders — with cyber guardrails off for that cohort so they can use its full defensive cybersecurity capabilities. Broader access for developers, enterprises, and consumers follows after more guardrail iteration, starting with paid API customers and Google AI Ultra subscribers. Google also says it is engaged in the U.S. government’s voluntary pre-release model access process.

DeepMind’s Fairwind page says partners can use Argon standalone or with CodeMender, and that the program already works with more than 650 partners globally.

Benchmarks cited

• DeepSWE v1.1: 77.9% (new SOTA for long-horizon software engineering).
• AutomationBench: 51.3%, #1 (Zapier end-to-end business-function benchmark).
• LVBench: 91.7% SOTA (long video understanding).
• CWE-bench v1: 68%, tied for first (vulnerability remediation).

Google also says Argon leads the Vals Index and leads on Vals Finance Agent v2 and Harvey’s Legal Agent Benchmark; those posts do not publish numeric scores for those three in the Argon announcement.

Internal Google examples

• Quantum: optimized a bottleneck subroutine 40% past the published baseline in minutes.
• Memory: agent fleet freed over 300 TiB once rolled out, with an estimated 500 TiB–1 PiB in total savings.
• Migrations: C/C++→Rust work from tens of thousands of lines (re2, libgav1) up to 800K+ lines for the Fuchsia Zircon kernel; for libgav1, agents replaced 32K lines of SIMD and produced a memory-safe decoder 2.7× faster than the prior Rust port with identical video output.

Wiz is already using Argon through its Scan for Good initiative; Google says the model found a critical healthcare-software exposure that prior frontier models missed.

Sources: Google blog (Koray Kavukcuoglu, Sep 30, 2026); DeepMind Fairwind Program.