Google has opened Gemini 4 Argon to developers through the paid Gemini API, moving the model beyond the limited rollout announced on September 30. The October 13 release also brings Argon to Google Antigravity’s desktop app and command-line tool for Google AI Ultra subscribers and eligible Gemini Enterprise customers. [1, 2, 3]
For developers working on a Mac, Antigravity provides a direct way to use the model in coding workflows. Google offers the desktop app for both Apple silicon and Intel Macs running macOS 12 or later. The API is also available through Google AI Studio; Argon is not included in the API’s free tier. [4, 5]
What the million-token limit means
Google documents a unified context window of 1,048,576 tokens, shared by the input, the model’s thinking and its answer. A large prompt therefore leaves less room for reasoning and output. [6, 7]
On the Interactions API, ordinary synchronous requests stop at 65,536 generated tokens. Going beyond that requires background execution and an explicitly higher output limit. Google’s servers continue the response in successive chunks. Background execution stores the interaction on Google’s servers so the client can retrieve results. Developers who manage continuation themselves can use the separate GenerateContent API flow. [8, 9]
That gives difficult coding or research tasks more space to finish, though Google says most coding and agent tasks fit within the normal 64K limit. Longer generation can take minutes, and a low output cap can exhaust the budget during thinking before a useful answer appears. [8]
Developers can select low, medium or high thinking levels. Raising the output ceiling gives the model more room; it does not force the model to use the full allowance. [7, 8]
The launch price has an end date
Google’s introductory Standard API prices end November 30, 2026. Output billing includes thinking tokens. [5]
| Token type | Through Nov. 30, 2026 | From Dec. 1, 2026 |
|---|---|---|
| Input | $2 | $4 |
| Output, including thinking | $10 | $20 |
Long-running requests also accumulate input charges as earlier material is fed into subsequent chunks. Google documents a 95% implicit-cache discount for repeated input in these long-decode steps. Cancelled or failed interactions still incur charges for completed work. [8]
Teams can now test Argon against their own code and workflows. Its longer generation path brings specific integration, storage and billing requirements alongside the larger token budget.
