Updated 11:36
Gemini 4 Argon reaches the Gemini API at $2 in, $10 out until November 30, and explicit caching doesn't come with it
Gemini 4 Argon's pricing, migration notes and model card, read together: what changes for existing Gemini 3.x code and what Google says about cyber risk.
When XT.PT last looked at Gemini 4 Argon, it was a model for Google's vetted Fairwind cyber defenders, with published prices and no date for anyone else. That changed on October 13, 2026. The Gemini API changelog entry for that day reads, in full on the point that matters, that gemini-4-argon is generally available as the most capable model for "complex coding, agentic, math, and science tasks." The same day, Google's consumer blog put Argon in the Gemini app for Google AI Pro and Ultra subscribers.
The developer documentation now fills in what the announcement left open, and some of it will break existing code.
The price, and the date it doubles
Google's pricing page lists an introductory rate through November 30, 2026 and a standard rate from December 1:
| Per 1M tokens | Through Nov 30 | From Dec 1 |
|---|---|---|
| Input | $2.00 | $4.00 |
| Output (incl. thinking) | $10.00 | $20.00 |
| Context caching | $0.10 | $0.20 |
| Batch input | $1.00 | $2.00 |
| Batch output | $5.00 | $10.00 |
Argon is not on the free tier. Search and Maps grounding get 5,000 free requests a month, shared across Gemini 3 and newer models, then $14 per 1,000. Anyone running a cost model on November's bills should budget for December at twice the rate.
What changes for existing code
The model page lists a 1,048,576-token context window that is "unified input and output token limit," and 65,536 output tokens by default for synchronous requests, extendable to the full window with what Google calls long decode. The "latest model" migration page spells out the behavior changes:
- Explicit caching is gone. "Explicit caching is not supported," says the model page. Requests that reference
cached_contentmust drop it. Implicit caching is on by default for prompts of at least 4,096 tokens, with a 95% discount on hits. - Sampling parameters are ignored.
temperature,top_pandtop_kshould be removed. For now the page says they are ignored, not rejected. - Thinking eats the output budget. Thinking and output share
max_output_tokens, so a low ceiling can truncate answers. Google recommends at least 32,768 for apps that used lower limits. Thinking levels arelow,medium(the default) andhigh. - Long output arrives in pieces. Outputs beyond 64K tokens require background execution and are generated in hops of up to 64K, each adding a
model_outputstep. The page warns that joining only one step truncates the response. Settingbackground: truewithoutmax_output_tokensstops at 64K. - Timeouts. Hard prompts can take several minutes; Google says to use background execution or set HTTP timeouts to at least 15 minutes.
Requests go through the Interactions API (client.interactions.create). Argon supports function calling, structured outputs, code execution, search grounding, file search, URL context, and computer use in preview. It does not do audio or image generation or the Live API.
The cyber question, answered in the model card
The general release is not the Fairwind configuration. Google's model card, published October 13, is direct about capability:
We cannot rule out that Gemini 4 Argon reaches the cybersecurity CCL.
Google DeepMind, Gemini 4 Argon model card
The stated basis is that "The model provides substantial uplift by our evaluations and quantitative modeling." Google's response, per the card, was a safety case for cyberoffensive capability and safeguards it assesses bring the risk below that Critical Capability Level. The card does not list the safeguards; it points to a separate Safety & Security Report. It also resolves a gap in the original announcement: the 68.0% CWE-bench v1 score is a tie with GPT-6 Astra.
On refusals, the card reports 9.60 percentage points fewer unjustified refusals than Gemini 3.1 Pro, while multilingual and image-plus-text safety scores got worse, by 3.16 and 9.29 points.
Why it matters
Argon is now a model you can call, at a price that holds for six weeks. Teams migrating from Gemini 3.x should treat the caching removal and the shared thinking budget as the two changes most likely to show up as silent cost increases or truncated answers, not as errors.
Primary sources: Gemini API changelog, Gemini 4 Argon model page, Latest Gemini models, Gemini API pricing, Gemini 4 Argon model card, Google: Gemini 4 availability, read 2026-10-20.