Gemini 4 Argon raises the output token ceiling to 1M

Google DeepMind announced Gemini 4 Argon, a frontier model built to hold deep reasoning across long horizon work. The headline number is an output token limit of 1 million, up from 64K, which is what makes very long generations practical rather than iterative.

Introductory pricing is $2 per million input tokens and $10 per million output, with cached input at 95% off. There is a footnote that matters: $4 and $20 apply after the introductory period. Read that before building a cost model on the launch price.

On benchmarks, Argon sets state of the art on DeepSWE v1.1 at 77.9%, leads the Vals Index, leads Vals Finance Agent v2 and Harvey’s Legal Agent Benchmark, ranks first on AutomationBench at 51.3%, posts state of the art LVBench at 91.7% for long video understanding, and ties for first on CWE-bench v1 at 68%.

The cyber posture is the unusual part

Argon is trained to autonomously find, validate and patch critical vulnerabilities, and it will be released without cyber guardrails to trusted defenders through the Fairwind Program and to internal Google teams. Guardrails against harmful cyber and CBRN requests still apply under the Frontier Safety Framework for everyone else. Google says it leads Gray Swan’s Indirect Prompt Injection benchmark and that the rollout follows the US government’s voluntary pre-release model access process.

The named partner is Wiz, using Argon through its Scan for Good initiative, where an early demo found a critical vulnerability exposing sensitive personal information across hospital software that earlier frontier models missed.

Google also reports internal results: a 40% improvement over a published baseline on quantum subroutine optimization measured in qubit times gate minutes, and a fleet of agents that freed more than 300 TiB of memory across Google data centers with an estimated 500 TiB to 1 PiB total savings. The most concrete one is a C and C++ to Rust migration on libgav1, where agents replaced 32,000 lines of SIMD code and produced a memory safe decoder running 2.7 times faster than the previous Rust port with identical video output.

Availability is phased, starting with trusted cyber defenders, then paid API customers and Google AI Ultra subscribers. It is not broadly available at announcement.

(Original source: Google DeepMind)