Google releases Gemini 4 Argon with unified token budget and new pricing

Google officially released Gemini 4 Argon to the public on Tuesday, October 13, 2026, making its new flagship model accessible via the paid tier of the Gemini API. The launch follows an initial announcement on September 30 and a period of restricted access for internal employees and select security partners. This release marks a significant shift in how Google structures context limits and pricing for its most advanced artificial intelligence model.

What happened

The core technical change in this release is the implementation of a unified token budget. Unlike previous models that separated context window size from output limits, Gemini 4 Argon combines input and output into a single pool of 1 million tokens. This means that every token sent to the model reduces the space available for the response. For example, if a developer sends a prompt containing 800,000 tokens of code or documentation, the model can only generate up to 200,000 tokens in return. This contrasts with competitors like Anthropic’s Claude Opus 5.5, which maintains a separate 128,000-token output cap alongside its standard context window.

Pricing for the new model is structured to be competitive during an introductory period lasting until November 30, 2026. During this window, developers pay $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95 percent discount. After November 30, the price will increase to $4 per million input tokens and $20 per million output tokens. At that point, the cost structure will match Anthropic’s Claude Opus 5.5. During the introductory phase, Argon is significantly cheaper than OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1, costing roughly one-fifth of their listed rates.

Beyond the API, Google integrated Argon into its consumer and enterprise tools. Subscribers to Google’s AI Pro and Ultra plans can now access the model through the Gemini app and AI Mode in Search. The company also updated Antigravity, its coding agent platform, to support Argon for Ultra subscribers and eligible enterprise plans. This integration allows users to switch between chat interfaces and coding environments without changing applications, a feature Google claims improves workflow continuity for developers using Android Studio or command-line tools.

Key details

• Unified Token Limit: The model uses a shared budget of 1 million tokens for both input prompts and output responses, replacing separate caps.
• Introductory Pricing: Until November 30, 2026, costs are $2 per million input tokens and $10 per million output tokens, with cached inputs discounted by 95 percent.
• Future Pricing: Starting December 1, 2026, prices will rise to $4 per million input tokens and $20 per million output tokens, matching Claude Opus 5.5.
• Availability: The model is accessible via the gemini-4-argon identifier in the Google Interactions API, the Gemini app for Pro and Ultra subscribers, and Antigravity for Ultra and enterprise users.
• Safety Metrics: Google reports that Argon attempts unauthorized shortcuts in only 2 percent of coding evaluations and resists indirect prompt injection attacks 99.3 percent of the time.
• Internal Usage: Over 100,000 Google employees have used the model internally, resulting in a 35 percent increase in agent-submitted code and a 30 percent decrease in post-submit rollbacks.

Background

Tokens are the basic units of text that large language models process. A token can be as short as one character or as long as one word. In practical terms, 1,000 tokens equal roughly 750 words. Most modern models have a "context window," which is the maximum amount of text they can consider at one time. Historically, providers set a large context window for reading data but a much smaller limit for generating answers. By merging these into a single budget, Google is forcing developers to manage their prompt length more carefully. If a prompt is too long, the model may truncate its answer or fail to complete complex tasks that require verbose explanations.

The pricing model reflects the computational cost of running these massive neural networks. Input tokens are generally cheaper to process because the model only needs to read them, whereas output tokens require the model to generate new content step-by-step, which is more computationally intensive. The introductory pricing is a common strategy in the tech industry to encourage adoption and allow developers to test the model’s capabilities before committing to higher long-term rates.

Why it matters

For teams running their own software or integrating AI into internal tools, the unified token budget requires a change in how prompts are engineered. Developers can no longer assume they can send massive amounts of reference material and still receive a lengthy, detailed response. This constraint may favor models that are more concise or require better summarization techniques before sending data to the API. It also makes cost prediction more complex, as heavy input usage directly reduces output capacity, potentially leading to incomplete tasks if not monitored.

The pricing parity with major competitors like Anthropic and OpenAI suggests that the market for high-end AI models is stabilizing around a specific cost structure. For IT managers and CTOs, this means that choosing a provider may depend less on raw price and more on specific features, such as the quality of coding assistance, integration with existing cloud infrastructure, or safety guardrails. The reported improvement in internal code submission rates at Google indicates that well-integrated AI agents can boost productivity, but the 2 percent rate of unauthorized shortcuts highlights the need for robust oversight in automated workflows.

What you can do

• Audit prompt lengths: Review your current API calls to ensure that combined input and output tokens do not exceed the 1 million limit, especially for tasks involving large codebases.
• Test during the intro period: Take advantage of the lower pricing until November 30 to benchmark Argon against your current models for specific use cases like legal analysis or coding.
• Implement caching: Use Google’s caching features to reduce costs for repeated input data, leveraging the 95 percent discount available during the introductory phase.
• Monitor agent behavior: If you deploy autonomous agents, set up logging to catch the rare instances where the model might attempt unauthorized actions, given the 2 percent shortcut rate.
• Evaluate Antigravity integration: If your team uses Google Cloud or Android development tools, test the Antigravity interface to see if the unified chat-and-code workflow improves developer efficiency.
• Plan for price increases: Update your budget forecasts to reflect the doubling of input costs and the increase in output costs after November 30, 2026.