Google Releases Gemini 4 Argon with 1M-Token Output Window and Frontier Benchmark Results

Google DeepMind has officially released Gemini 4 Argon, positioning it as a new era of frontier intelligence designed to solve complex engineering and scientific problems. The model is now accessible via the Gemini app and Google Antigravity for Google AI Pro and Ultra subscribers.

Key Performance Metrics

Gemini 4 Argon demonstrates leading performance across multiple benchmarks, particularly in knowledge work, agentic coding, and long-context reasoning:

• Knowledge Work: Achieved the highest score on Vals Index (68.9%) and AutomationBench (51.3%), surpassing GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5.
• Agentic Coding: Recorded a top score of 77.9% on DeepSWE v1.1 and 91.9% on Vibe Code Bench.
• Long Context: In GraphWalks benchmarks, it achieved 99.7% F1 score for BFS up to 128k tokens and 84.2% for 256k to 1M tokens, significantly outperforming competitors in extended context retention.
• Science & Math: Top scores on LABBench 2 (88.8%) and RiemannBench (76.0%).

1M-Token Output Window

A defining feature of Gemini 4 Argon is its 1M-token output window, enabling single-pass generation of massive artifacts without truncation. This capability supports:

• Complex Code Generation: Full multi-file application scaffolding and extensive codebase refactoring in one go.
• Scientific Derivations: Completion of 100K+ token reasoning trajectories, including step-by-step mathematical proofs and quantum error correction decoders.
• Legacy Translation: End-to-end translation of monolithic legacy files and complex database schemas while preserving inter-file dependencies.

Real-World Applications and Partners

The release highlights significant efficiency gains in enterprise and research workflows:

• Chrome Team: Assisted in refactoring a 6 million-line legacy folder within Chromium’s 40-million-line codebase, reducing a months-long task to weeks.
• xTB (Quantum Chemistry): Improved molecular structure prediction speed by 2.3x and conformer search by 1.8x using Google Antigravity.
• Google ADK: Reduced in-memory session preparation time by 40% and accelerated cold imports by 21%.
• Gemma 4 E2B Optimization: Achieved a 2.85x improvement in end-to-end inference speed on A100 GPUs (from ~37,900 to 108,000+ tokens/sec) with zero quality degradation.
• Isomorphic Labs: Shortened complex PDF and patent synthesis processes from weeks to hours.
• Box Evaluation: On over 200 complex synthetic longitudinal charts, Gemini 4 achieved the highest recall and ICD-10 code specificity among the models evaluated, producing concise problem lists without external lookups.

Safety and Research

Gemini 4 Argon is evaluated as the most resilient model against indirect prompt injections. It adheres to the Frontier Safety Framework, allowing legitimate dual-use scientific research while blocking harmful requests. In cybersecurity, it tied for first place on CWE-Bench (68%) and is being used in the Wiz 'Scan for Good' initiative to proactively protect medical infrastructure. Research partnerships have yielded new counterexamples to 'Grad's conjecture' in fusion physics and identified novel enzyme designs with confirmed catalytic activity.