Google has made its flagship artificial intelligence model, Gemini 4 Argon, available to developers subscribed to the paid Gemini API plan, following initial availability limited to two trusted testers and the Fairwind program for cyber defense teams. The model can be accessed through Google AI Studio and is also available in a number of Google tools for developers and users.
Introductory Price and Unified Token Limit
Through November 30, using Gemini 4 Argon costs $2 per million input tokens and $10 per million output tokens. After the introductory period ends, the price will rise to $4 for input and $20 for output per million tokens, with a 95% discount on cached input tokens during the introductory period.
The model is invoked using the identifier gemini-4-argon through the Interactions API, with support for Google GenAI SDK packages. Android developers can also integrate it into their applications through Firebase AI Logic.
One of the release’s most notable features is that the one-million-token limit is no longer a separate ceiling for output; instead, it is a shared budget between what enters the model and what it produces. If a developer sends 800,000 tokens of code or context, only 200,000 tokens will remain for the response within that same limit.
What Changes for Developers in Practice?
The large budget gives the model room to work on multi-step software and scientific tasks, in addition to agentic tasks that require long context. However, the unified limit requires teams to measure input size before relying on long responses; increasing the context does not come without a cost in terms of available output space.
Google says Argon ranked first on the Vals benchmark, which measures model performance in finance, programming, law, and taxes. The company also reported that it scored 87% on the ERQA test for physical and spatial reasoning, compared with 82.3% for Claude Opus 5.5, 81% for GPT-6 Astra, and 74% for Claude Fable 5.1.
However, the source material indicates that comprehensive independent comparisons remain limited and that some previous tests do not guarantee that the model evaluated matched the version currently available. Bloomberg also reported observations about difficulties with some programming tasks, particularly front-end design, which Google denied, calling the characterization inaccurate.
Availability in Programming Tools and Safety
Argon is also available on the Antigravity platform for running coding agents through the editor, terminal, and browser, although this is limited to Google AI Ultra subscribers and certain Gemini Enterprise plans. It also appears in the Gemini app for AI Pro and Ultra subscribers and in AI Mode in Search, while Google is working to integrate the API into the Gemini app for macOS.
Google says that more than 100,000 of its employees use Argon monthly through an internal version of Antigravity, and that code delivered by agents increased by 35% while post-integration rollback operations fell by 30% over four weeks. These are internal company figures, not an independent evaluation.
On safety, Google says the model is trained to request human approval before carrying out irreversible actions and to report obstacles rather than work around them. It said that attempts at unauthorized shortcuts appeared in 2% of programming tests, while indirect prompt-injection attacks succeeded in 0.7% of attempts in the Gray Swan test, which consisted of 15 adaptive attempts.
Why Does This Release Matter?
The development is not limited to increasing context size; combining input and output into a single budget changes how developers design long-running tasks and coding agents. Users will need to balance the size of files, instructions, and conversation history against the required result length, while continuing to monitor exploit rates and unauthorized shortcuts because the reported percentages are not zero.