Google is making Gemini 4 Argon globally available with API prices set at half its announced standard rates until November 30, 2026. Developers can call the model as gemini-4-argon, paying $2 per million input tokens and $10 per million output tokens during the introductory period.
The October 13 availability announcement also brings Argon to paying users across Gemini, AI Mode in Search, and Antigravity, but those products don't share the same subscription requirements. Google AI Pro qualifies for the Gemini app and AI Mode; Antigravity access is specified for Google AI Ultra subscribers.
For teams evaluating the release, two details deserve particular attention: the API rates double after the introductory period, and Argon's 1-million-token context window is a combined input-and-output budget, not a million tokens of input plus another million for the response.
Where Gemini 4 Argon Is Available
This is a broad availability expansion rather than Argon's first announcement. Google introduced the model on September 30, initially rolling it out to trusted cybersecurity defenders through its Fairwind Program while preparing wider access.
The October 13 release opens several distinct routes:
| Product or access route | Announced access |
|---|---|
| Gemini API through Google AI Studio and Gemini platform | Developers and enterprises, using model ID gemini-4-argon |
| Gemini app | Google AI Pro and Google AI Ultra subscribers |
| AI Mode in Search | Google AI Pro and Google AI Ultra subscribers |
| Google Antigravity | Google AI Ultra subscribers |
| Gemini for business | Available on web, desktop, and mobile, according to Google's Cloud announcement |
The subscription distinction matters most for people choosing between Pro and Ultra. Google's announcement includes both tiers for the Gemini app and AI Mode, but names only Ultra for Antigravity. A Pro subscription therefore shouldn't be treated as evidence of Argon access inside Antigravity.
Enterprise access also needs to be considered separately from consumer subscriptions. Google lists Gemini for business as an available product surface, alongside API access for builders. That doesn't establish that every enterprise contract has identical entitlements, usage allowances, or deployment settings.
Likewise, Google's description of the rollout as “globally available” should not be read as a promise of unrestricted access for every account. The announcement establishes broad distribution across these products, with product and subscription conditions still applying. It does not provide a country-by-country access matrix or a universal quota.
For an application integration, the actionable identifier is gemini-4-argon. The product name alone isn't the model identifier to use in API configuration.
The Introductory API Price Ends After November 30
Google's Cloud release announcement specifies both the temporary and standard token rates:
| API token category | Through November 30, 2026 | After November 30, 2026 |
|---|---|---|
| Input | $2 per million tokens | $4 per million tokens |
| Output | $10 per million tokens | $20 per million tokens |
| Cached input | 95% below the input rate | 95% below the input rate |
Applying the stated caching discount produces cached-input token rates of $0.10 per million during the introductory period and $0.20 per million afterward. Those two figures are calculations from Google's published discount, rather than separately quoted prices.
The change means ordinary input and output rates double on December 1. Teams evaluating Argon in October or November should therefore calculate production economics at the standard rate, even if their initial experiments cost less.
For example, consider an illustrative request with 800,000 uncached input tokens and 100,000 billed output tokens:
- At introductory rates, input costs $1.60 and output costs $1, totaling $2.60.
- At standard rates, input costs $3.20 and output costs $2, totaling $5.20.
That calculation covers only the stated token charges. It is not a complete estimate for an agent that makes repeated model calls or uses other paid services.
Caching could materially change the economics of workflows that repeatedly use the same large body of context. But the 95% reduction applies to cached input tokens, not all input and output automatically. A realistic estimate should distinguish reusable cached material, new input, and generated output rather than discounting the whole workload.
The introductory offer is useful for evaluation. It is a weaker foundation for a long-term business case unless the application also works at December's prices.
What the 1M-Token Context Actually Allows
The Gemini 4 Argon model card describes support for text, image, audio, and video inputs within a combined input-output context window of up to 1 million tokens. Its listed output modality is text.
“Combined” is the operative word. Instructions, supplied material, conversation context, and the generated response must fit within the overall allowance. Filling the window with input leaves less room for output.
The earlier September announcement emphasized expanding the output token limit to 1 million tokens. Read alongside the model card, that should not be interpreted as permission to send a million-token input and then generate an additional million-token answer. Developers should budget against the shared limit and check the applicable endpoint documentation before configuring maximum request sizes.
In practice, a larger window gives teams more room to keep related evidence together. A code-analysis workflow could include source files, design notes, and screenshots in one request. A document-analysis workflow could combine text with charts or scanned material rather than relying entirely on a text-only extraction step.
Those are potential workflow benefits, not guarantees that every large request will be handled accurately. The window specifies capacity. It does not establish perfect recall, correct interpretation of every image, or reliable reasoning across every connection in a long record.
There is also no single dependable conversion from 1 million multimodal tokens into a fixed number of pages, images, or hours of video. The usable amount depends on the material and how it is represented.
Google's own long-context results illustrate the distinction between capacity and performance. In its model card, Argon scores 99.7% F1 on the GraphWalks breadth-first-search evaluation at lengths up to 128,000 tokens, compared with 84.2% F1 in the 256,000-to-1-million-token range. These are first-party reported benchmark results, and the task is not a substitute for testing a real document collection. They nevertheless caution against assuming that a larger supported window produces equally strong results at every length.
Google's Rankings Are Claims, Not Independent Validation
Google says Argon ranks first on the Vals Index, which covers work across domains including finance, coding, legal, and tax. Its model card reports a 68.9% Vals Index result.
That ranking should be described as Google's reported result, not as independently verified leadership established by this article. The supplied evidence consists of Google's announcement, Cloud post, and model card, rather than a separately checked benchmark-provider report.
Sources
- October 13 availability announcementblog.google
- introduced the model on September 30blog.google
- Cloud release announcementcloud.google.com
- Gemini 4 Argon model carddeepmind.google
