Google Launches Gemini 4 Argon: Targets Complex Reasoning Tasks with Introductory API Pricing at $2 Input and $10 Output per Million Tokens

Google Releases Gemini 4 Argon: Available Today for Pro and Ultra Members; API Pricing at $2/M Input Tokens,$10/M Output Tokens

On October 13, Google released Gemini 4 Argon, the first model in the Gemini 4 series. It is designed for complex, multi-step tasks such as coding, finance, law, cybersecurity, and scientific research. Google stated that other models in the series will be released later.

[Who Can Use It and Where]

• General Users: Subscribers to Google AI Pro or Ultra can access it today via the Gemini App and AI Mode in Google Search. Free-tier users do not have access yet.
• Developers and Enterprises: Access is available through the paid tier of the Gemini API, with entry points including AI Studio, the Gemini platform on Google Cloud, and Android Studio. The free API tier does not include this model.
• Antigravity: Google’s AI coding tool, Antigravity, now includes Argon. Individual users require an Ultra subscription to access it.

Additionally, there are two application channels:

1. Researchers can apply for a six-month free trial of the highest usage tier of Antigravity (limited to 10,000 slots, subject to eligibility and capacity screening).
2. Cybersecurity defense teams can apply to join Google’s Fairwind program to receive a version with cybersecurity restrictions removed.

[API Pricing]

Pricing is based per million tokens (the basic unit of text processing for the model). Current introductory pricing is:

• Input: $2 per million tokens.
• Output: $10 per million tokens (includes tokens consumed during model reasoning/thinking).
• Cached Input: Repeated input hits cost only 5% of the standard rate ($0.10 per million tokens).

Introductory pricing ends on November 30. Starting December 1, prices will double:

• Input: $4 per million tokens.
• Output: $20 per million tokens.
• Cached Input: $0.20 per million tokens.

For non-time-sensitive tasks, Batch or Flex tiers are available at half the standard price. The official pricing page does not offer tiered pricing based on prompt length; longer inputs remain at the same unit price.

[What Does the Unified 1 Million Token Limit Mean?]

Argon has a context window of 1,048,576 tokens, shared across input, reasoning, and output. For example, if you input 600,000 tokens of code and documentation, the model can generate approximately 440,000 additional tokens, with reasoning tokens counted within this limit.

Standard calls have a maximum output of 65,536 tokens per request. To generate longer outputs, developers must use "Long Output" mode as described in the developer documentation: requests are set as background tasks, where the server generates segments of approximately 64,000 tokens each, continuing from the previous segment until completion or until the user-defined limit is reached. A single task can run for up to 24 hours. This feature is intended for long-running autonomous reasoning and batch code generation, not typical chat interactions.

The documentation notes that during segmented generation, each segment re-processes prior content as part of the input. Consequently, reported input token usage may significantly exceed the actual content size. The billing implications of this overhead are not specified in the documentation.

[Comparison with Competitors]

Independently verifiable results come from third-party evaluation firm Vals AI’s Vals Index, which measures performance in areas like finance, coding, law, and tax. On the leaderboard updated October 12, Argon ranks first with 68.90%, followed by Anthropic’s Claude Sonnet 5.5 (67.04%) and Claude Opus 5.5 (66.97%), leading by less than two percentage points.

Other metrics are self-reported by Google:

• ERQA Benchmark (Physics and Spatial Understanding): Argon scored 87%, which Google claims exceeds all other frontier models.
• Gray Swan Prompt Injection Test: In tests where malicious instructions are hidden in web pages or emails to induce agents to perform unintended actions, the success rate of targeted attacks dropped to 0.7%.

Regarding everyday usability, Google states that Argon provides more direct answers, reduces ambiguity and repetition, and adheres better to style guidelines. These comparisons are against previous versions of Gemini.

[Real-World Performance]

The following cases are provided by Google.

• Agent Development Kit (ADK): Google’s open-source agent development framework contains approximately 200,000 lines of Python code. The team provided Argon with project rule files and a one-sentence goal: reduce framework overhead without breaking existing applications. Argon ran overnight autonomously, identifying optimizations that have since been merged into the codebase. Session data preparation time under long sessions decreased by 40%, and initial import speed improved by 21%.
• Chromium Refactoring: The Chrome team used Argon to refactor legacy Chromium code. Chromium comprises 40 million lines of code, including a legacy folder of about 6 million lines mixing various browser features. Argon helped the team extract core web processing logic, involving changes to hundreds of files. Google reports that work estimated to take months was completed in weeks. According to Google, this lays the groundwork for moving web content between tabs, sidebars, and floating windows without reloading, while also eliminating an entire class of crashes.
• Internal Engineering Metrics: Google reports that internal engineers using Argon within Antigravity saw a greater than 35% increase in agent-submitted code volume and nearly a one-third reduction in code rollbacks.

↳ Quote @sundarpichai (Sundar Pichai): Gemini 4 Argon