Google Launches Gemini 4 Argon Model with Million-Token Output Support and 80% Off Launch API Pricing

On October 14, 2026, Beijing Time, Google, under Alphabet, officially launched its cutting-edge large language model, Gemini 4 Argon. Starting today, the model is available via the Application Programming Interface (API) on Google AI Studio and the Gemini platform for developers and enterprise customers worldwide, and is also integrated into the Google Antigravity workflow platform for Google AI Ultra subscribers, as well as the search AI mode (AI Mode) for Google AI Ultra and Pro subscribers.

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

Unified Pricing and Million-Token Generation Window

In terms of billing structure and system specifications, Gemini 4 Argon sets a unified limit of 1 million input/output Tokens. Against the backdrop where traditional large models generally restrict single-generation length to 4,000 to 64,000 Tokens, Argon allows agents to complete a single continuous output of up to 1 million Tokens in one session, resolving the issue of long-horizon agents interrupting operations due to output truncation limits during large codebase refactoring or complex document reasoning.

Regarding invocation costs, the newly introduced rate of $2 per 1 million inputs and $10 per 1 million outputs represents an introductory promotional price. Compared to similarly tiered frontier models on the market, such as GPT-6 Astra priced at $10 per 1 million inputs and $50 per 1 million outputs, Argon's introductory rate is only one-fifth (effectively reducing invocation costs by 80%). Under the context caching mode, the cost per 1 million input Tokens drops to $0.10. When agents frequently retrieve and invoke the same 1-million-Token codebase or legal dossier in long-text environments, the marginal input cost per interaction shrinks from $2.00 to $0.10, saving 95% on redundant data parsing overhead. After the promotional period ends, the model's standard pricing will revert to $4 per 1 million inputs and $20 per 1 million outputs.

Core Economic Benchmark Vals Index and Professional Benchmark Evaluations

According to Google DeepMind's published technical report, Argon primarily optimizes its architecture for multi-step, long-chain reasoning tasks.

In the Vals Index benchmark evaluating key economic productivity performance across finance, code, law, and taxation, Gemini 4 Argon achieved an overall score of 68.9%, ranking first among mainstream frontier models, surpassing the industry average of 63.1% to 67.0% in comparable tests.
The model also released specific evaluation results for other standardized professional task tests:

• In the Zapier AutomationBench test measuring end-to-end workflow execution capabilities across business functions, it led with a 51.3% completion rate;
• In the LVBench test assessing long-video and multimodal spatiotemporal understanding, it achieved a 91.7% accuracy rate;
• In the CWE-bench v1 benchmark targeting cybersecurity vulnerability localization and automated remediation, it tied for first place with a 68.0% repair success rate.

For system security during long-horizon automation execution, Google has integrated multiple layers of dynamic security defense mechanisms at the base level to prevent the model from encountering prompt injection and jailbreak commands when invoking external terminals and code sandboxes.

Complex Engineering Deployment and Codebase Refactoring Performance

During internal testing and early enterprise customer trials at Google, Argon was deployed in real-world engineering maintenance scenarios. In a test for refactoring the underlying architecture of the Chromium browser, the engineering team used Argon to decouple the Chromium project containing 40 million lines of code, successfully completing an architectural rewrite that separated approximately 6 million lines of legacy web processing functionality from the core, thereby enabling seamless rendering across multiple tabs and floating windows.

In scientific computing and model inference optimization, the research team utilized Argon to optimize code for the quantum chemistry simulation package xTB, increasing its molecular 3D spatial configuration prediction speed by 2.3 times and conformational search efficiency by 1.8 times. When deploying the open-source model Gemma 4 E2B using vLLM, Argon automatically adjusted attention mechanisms and speculative decoding paths, boosting end-to-end inference throughput in an A100 hardware environment from approximately 37,900 Tokens per second to over 108,000 Tokens per second, achieving a 2.85x inference acceleration.

For external enterprise integration, Deloitte and biomedical research organization Axiom Bio have already connected the model to their long-cycle professional reasoning pipelines. According to Google's official schedule, the introductory promotional rates are planned to run until the end of Q4 2026.

Related Reading:

Google Launches Three Models Including Gemini 3.6 Flash: Saves 17% Token Costs

Google Introduces Agent Gemini Spark for AI Pro Users Outside the US

Join the official Coincamps community:

X: https://x.com/coincamps

Telegram: https://t.me/coin_camps