Gemini 4 Argon Launches with 1M Output Tokens, Leads Coding Benchmarks

Google officially released Gemini 4 Argon globally at 00:30 UTC, making it immediately available through the Gemini API, the Gemini App for Pro and Ultra subscribers, and the Antigravity CLI for Ultra subscribers. The author upgraded to the Ultra tier and confirmed the model appears as "Gemini 4 Argon (Medium)" in the Antigravity model selector.

Key Specification: 1 Million Token Output Limit

The most striking change is the maximum output token limit, raised from 64K in the previous generation to 1,000,000 tokens — a 15× increase. While many models advertise million-token context windows for input, they often cap output at 32K or 64K, causing truncation during long code generation or extended reasoning chains. A 1M output ceiling gives the model ample space to complete cross-file refactors or generate lengthy technical documents in a single pass.

Internal Engineering Case Studies

Google's technical blog highlights three production deployments inside the company:

Large-scale C/C++ to Rust migration: Argon agents are migrating core libraries — from the 10K-line re2 regex library and the open-source libgav1 video decoder up to the 800K-line Fuchsia Zircon kernel. In libgav1, Argon ran multiple profile-guided optimization rounds, replacing 32K lines of hand-written SIMD with safe Rust that the compiler can auto-vectorize. The resulting memory-safe decoder runs 2.7× faster than the prior Rust port and matches the highly optimized C++ original.

Data-center memory optimization: Argon agents autonomously analyzed global server fleet telemetry, identified and landed memory-saving patches, freeing over 300 TiB of physical RAM immediately; projected full rollout savings reach 500 TiB–1 PiB.

Quantum algorithm optimization: Argon helped researchers optimize a core subroutine's space-time volume (qubit count × gate count), improving the published baseline by 40% in minutes.

Benchmark Results

Google published head-to-head scores against GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5:

DeepSWE v1.1 (long-horizon software engineering): Gemini 4 Argon 77.9% vs. Claude Opus 5.5 74.2% vs. GPT-6 Astra 74.1%.

Vibe Code Bench: Argon 91.9%, leading the field.

GraphWalks (256K–1M context): Argon 84.2%, ahead of GPT-6 Astra (71.8%) by 12.4 points and Claude Opus 5.5 (66.8%) by 17.4 points.

Pricing

Introductory pricing: $2 per million input tokens, $10 per million output tokens, with cached input at $0.10 per million tokens (50% off). After the promotional period, standard pricing will be $4 input / $20 output per million tokens. Existing Google AI Pro, Ultra, or Antigravity users can switch to the model immediately.

Author's Note

The article itself was written by Gemini 4 Argon. The author, a nine-year AI algorithm veteran, plans deeper hands-on testing the following morning.