Google Launches Gemini 4 Argon Model with 1M Token Output and 80% API Price Reduction

Author: Shenchao TechFlow

On October 14, 2026, Beijing Time, Alphabet’s Google announced the official launch of the advanced large language model Gemini 4 Argon. The model is now available globally to developers and enterprise customers via the Google AI Studio and Gemini platform APIs, and is integrated into the Google Antigravity workflow platform for Google AI Ultra subscribers, as well as the Search AI Mode for Google AI Ultra and Pro subscribers.

During the model's initial launch phase, API pricing is set at $2 per 1 million input tokens and $10 per 1 million output tokens, with a 95% discount offered on context caching.

Uniform Pricing and Million Token Generation Window

In terms of billing structure and system specifications, Gemini 4 Argon sets a limit of 1 million unified input-output tokens. Against the backdrop of traditional large models typically restricting single-generation lengths to 4,000 to 64,000 tokens, Argon enables agents to produce a single continuous output of up to 1 million tokens per session, resolving the issue of interrupted operations in long-range agents during tasks such as large-scale codebase refactoring or complex document derivation due to output truncation.

On calling costs, Google’s introductory pricing for this launch is $2 per 1 million input tokens and $10 per 1 million output tokens. Compared to market-leading models such as GPT-6 Astra, which charges $10 per 1 million input tokens and $50 per 1 million output tokens, Argon’s launch rate is just one-fifth—equivalent to an 80% reduction in calling expenses. Under context caching mode, the cost per 1 million input tokens drops to $0.10. When agents repeatedly reference the same 1-million-token codebase or legal archive in long-text environments, the marginal input cost per interaction falls from $2.00 to $0.10, reducing redundant data parsing overhead by 95%. After the promotional period ends, the model’s standard pricing will revert to $4 per 1 million input tokens and $20 per 1 million output tokens.

Core Economic Benchmark: Vals Index and Professional Benchmark Evaluation

According to the technical report released by Google DeepMind, Argon is primarily optimized for multi-step long-chain reasoning tasks.

In the Vals Index benchmark evaluating key economic productivity metrics such as finance, code, law, and taxation, Gemini 4 Argon achieved a composite score of 68.9%, ranking first among mainstream state-of-the-art models and surpassing the industry range of 63.1% to 67.0% in similar tests.
In other standardized professional task tests, the model also published specific evaluation data:

Ranked first in the Zapier AutomationBench test measuring end-to-end workflow execution across business functions, with a completion rate of 51.3%;
Achieved 91.7% accuracy on the LVBench benchmark for evaluating long-video and multimodal spatiotemporal understanding;
Tied for first place with a 68.0% success rate in the CWE-bench v1 benchmark for locating and automatically fixing cybersecurity vulnerabilities.

To ensure system security during long-range automated execution, Google has integrated multiple layers of dynamic security defenses at the core level to prevent prompt injection and jailbreak commands when the model calls external endpoints and code sandboxes.

Implementation of complex engineering and codebase refactoring performance

Argon was deployed in real-world engineering maintenance scenarios during internal testing at Google and with early enterprise customers. During testing of the Chromium browser’s underlying architecture refactoring, the engineering team used Argon to decouple the Chromium project, which contains 40 million lines of code, successfully rewriting approximately 6 million lines of legacy web rendering functionality from the core, enabling seamless rendering across multiple tabs and floating windows.

In the area of scientific computing and model inference optimization, the research team used Argon to optimize the quantum chemistry simulation package xTB, achieving a 2.3x speedup in molecular 3D conformation prediction and a 1.8x improvement in conformation search efficiency. When deploying the open-source model Gemma 4 E2B using vLLM, Argon automatically tuned attention mechanisms and speculative decoding pathways, increasing end-to-end inference throughput on A100 hardware from approximately 37,900 tokens per second to over 108,000 tokens per second, resulting in a 2.85x acceleration in inference.

In terms of external enterprise integration, Deloitte has integrated this model into its long-cycle professional inference pipeline with biopharmaceutical R&D firm Axiom Bio. According to Google’s official schedule, the initial promotional rate is planned to run through the end of the fourth quarter of 2026.