Gemini 4 Argon Is Generally Available: API Price, Limits and What to Test First

The Gemini 4 Argon API went generally available on October 13, 2026, under the model ID gemini-4-argon. A week ago, only members of Google’s Fairwind security program could use it. Now any paid Gemini API project can call it, and Google AI Pro and Ultra subscribers can pick it in the Gemini app and in AI Mode in Search.

This piece follows our October 7 explainer on what Gemini 4 Argon is, which said Argon was Fairwind-only with no model ID. It covers the GA facts, the confirmed price and when the intro rate ends, the token limits, and where Argon runs. It also covers two quieter Gemini API changes from October 8 that can move your bill, and a short test plan against Claude Opus 5.5 and GPT-6.x. Status is as of October 14, 2026.

The short version: the intro price of $2 per million input tokens and $10 output ends on November 30. From December 1 it doubles to $4 and $20. The 1M-token output limit is real but needs background execution; synchronous calls stop at 65,536 tokens by default. Argon ignores temperature and does not support explicit caching. Price any test at the December rate, and check your Flash model IDs this week.

What changed on October 13

Google’s Gemini API changelog lists Gemini 4 Argon (gemini-4-argon) as generally available on October 13. The models page lists it under Stable with a "New Stable" tag. Google calls it its flagship reasoning model.

Google’s launch post, by Koray Kavukcuoglu, says the model is "now globally available today." That closes most of the open questions from our first post.

Gemini 4 Argon API pricing and the November 30 change

The pricing page lists two rates for Gemini 4 Argon API calls. Through November 30, 2026, standard input costs $2.00 per million tokens and output costs $10.00, thinking included. From December 1, those become $4.00 and $20.00. Cached input is $0.10, then $0.20.

There is no higher rate for long prompts, unlike the 3.1 Pro preview. There is no free tier. The pricing page also says Argon is not available on the Gemini API’s Google AI Pro plan. Search grounding gets 5,000 free requests a month, shared with other Gemini 3 and newer models, then $14 per 1,000.

The intro window is short. A project started now gets about six weeks at the lower rate. Any client estimate that runs past November should use the December figures.

Context, output and the parameters that changed

The Argon model page gives a 1,048,576-token window. It takes text, images, video, audio and PDF, and returns text. Thinking runs at low, medium or high, with medium as the default.

The 1M output figure needs a careful read. The default output limit is 65,536 tokens on a normal synchronous call. Longer answers need background execution with a higher max_output_tokens. Input, thinking and output share one window, so a long prompt leaves less room for output.

Google’s latest model guide lists changes that affect existing code:

• Sampling parameters are ignored. temperature, top_p and top_k are removed. Requests that send them still run.
• No explicit caching. Remove cached_content references. Implicit caching is on by default for prompts of at least 4,096 tokens.
• Long timeouts. Google says hard prompts can take several minutes and suggests HTTP timeouts of at least 15 minutes.

Argon supports function calling, structured outputs, code execution, search grounding and URL context. Computer use is in preview. It does not support the Live API or image generation.

Where the Gemini 4 Argon API and apps are available

Google’s launch post and developer post name these surfaces:

• Gemini API: paid tier, in Google AI Studio and Android Studio.
• Google Cloud: the launch post links Argon in Google Cloud’s Agent Platform console, formerly Vertex AI, and Cloud has a developer guide for it. We could not confirm the Cloud launch stage or regions.
• Gemini app: the Gemini app release notes say Argon is “rolling out globally for Google AI Pro & Ultra subscribers.”
• AI Mode in Search: Pro and Ultra subscribers, per the launch post. Search Engine Roundtable reported the same on October 13, citing Google’s Nick Fox and Robby Stein.
• Antigravity: Google AI Ultra subscribers.

Free Gemini app users do not get Argon. Google’s help page on model access says that since October 9, users without an AI plan default to Auto, and "most prompts will go to Flash-Lite." Staff on free personal accounts will not see Argon answers.

Workspace and Gemini Enterprise are not named. The launch post says businesses can try Argon in "Gemini for business," without saying which plans.

Two quieter October 8 changes that can raise your bill

The same changelog has two October 8 entries that matter more for most client sites than Argon does.

Flash IDs now route to newer models. Requests to gemini-3.7-flash now go to gemini-3.8-flash. Requests to gemini-3.5-flash now go to gemini-3.6-flash. The deprecations page says these endpoints are not shutting down. Your code keeps working, but a different model answers.

That can change cost without a code change. Our test of what Gemini 3.8 Flash costs per task found it costs 57% more to run than 3.7 Flash at the same rate card, because it writes more. Google does not say which rate routed gemini-3.5-flash calls bill at. Check your billing report for October 8 onward.

One Deep Research agent shuts down October 23. deep-research-pro-preview-12-2025 stops working that day. Google names deep-research-preview-04-2026 for speed and deep-research-max-preview-04-2026 for depth. The change is the agent value in interactions.create requests.

Our list of Gemini shutdown dates this fall has a search pattern you can extend with these IDs.

A short Gemini 4 Argon API test plan for agencies

The Gemini 4 Argon API now sits beside models agencies already use. Current list prices per million tokens, from each vendor’s pricing page on October 14:

From December 1, Argon costs the same per token as Claude Opus 5.5. Per token says little on its own. A plan that fits in a week:

1. Pick 30 to 50 real tasks from one client job: a migration script, a long report, a content audit.
2. Run each model at its default settings and the same output cap. Argon defaults to medium thinking.
3. Count cost per finished task, priced at Argon’s December rate. Include retries.
4. Time each run. Argon can take minutes on hard prompts, which matters for anything a visitor waits on.
5. Test one long job that needs more than 64K output. That is the one thing Argon does that the others cannot.

Our post on OpenAI’s recommended migration path shows why the cheaper model in a family is worth a run too. The same goes for Claude Haiku 5.5 on tagging work. ChatGPT users are on GPT-6 models in the Chat tab now, so staff may compare Argon with that too. Google’s benchmark claims, such as first place on the Vals Index, are a reason to test, not a result.

Frequently Asked Questions