No. Gemini 4 Argon is not on the Google API free tier, and Google states it plainly on its own pricing page: "Gemini 4 Argon is not available on the Free tier or on the Google AI Pro plan." The Gemini app does include Argon, but only for Google AI Pro and Ultra subscribers in selected countries, and the Antigravity path needs Ultra.
There is exactly one route that costs a new user nothing, and it is a Google Cloud credit rather than a free tier. Everything else on this page is about making paid usage go further, which is a different and more useful question than it first looks.
Every figure below comes from Google's own pages, read on October 15, 2026. No third-party claim is used, because the free-access question is exactly the one where a wrong blog is most expensive.
Two rows in that table are the usual source of confusion. The consumer plan and the API tier are separate products with separate rules: a Google AI Pro subscription includes Argon inside the Gemini app, and the same subscription does not provide Argon on the API. And Google AI Plus sits below Pro on the plans page without Argon listed among its benefits.
Google Cloud's free page states that new customers get $300 in free credit to try products, and that you are not charged until you activate a full paid account. Argon runs on the Agent Platform and Vertex AI side of Google Cloud, so within that credit the model is usable at no cost.
The sequence:
# 1. create a project and link the new-customer credit gcloud projects create argon-free-trial --name="Argon trial" # 2. enable the API you will call gcloud services enable generativelanguage.googleapis.com --project argon-free-trial # 3. authenticate for the enterprise path gcloud auth application-default login
Then make the first call against the model, using the enterprise client form Google's model guide documents:
# pip install --upgrade "google-genai>=2.30.1" from google import genai from google.genai import types client = genai.Client( enterprise=True, project="argon-free-trial", location="global", # global, us, or eu ) response = client.models.generate_content( model="gemini-4-argon", contents="Name three ways to check whether a credit is still active.", config=types.GenerateContentConfig( thinking_config=types.ThinkingConfig(thinking_level="LOW"), ), ) print(response.text)
`thinking_level="LOW"` is deliberate for a trial. Thinking tokens are billed as output, and output is the expensive side at $10 per million tokens during the introductory period, so a low thinking level on exploratory calls is the cheapest way to learn what the model does.
Set the thinking level to LOW and the request small, then measure what it actually consumed rather than guessing:
usage = response.usage_metadata print("prompt tokens:", usage.prompt_token_count) print("output tokens:", usage.candidates_token_count)
On the Agent Platform the documented field to watch is `initial_prompt_token_count` rather than `prompt_token_count`, because continuation requests re-prefill earlier partial output as input tokens and make the plain prompt count grow cumulatively across a long decode.
The most expensive mistake in this area is not overspending, it is building on a key that cannot reach the model. Check the tier first, with a request that either works or returns a body you can read.
export GEMINI_API_KEY="your-key" curl -sS -X POST \ "https://generativelanguage.googleapis.com/v1beta/models/gemini-4-argon:generateContent" \ -H "Content-Type: application/json" \ -H "x-goog-api-key: $GEMINI_API_KEY" \ -d '{"contents":[{"parts":[{"text":"reply with the single word: ok"}]}]}' \ -w '\nHTTP %{http_code}\n'
If your project has no credential attached at all, this is the response, captured against the live endpoint on October 15, 2026:
{ "error": { "code": 403, "message": "Method doesn't allow unregistered callers (callers without established identity). Please use API Key or other form of API consumer identity to call this API.", "status": "PERMISSION_DENIED" } }
If a key is attached and rejected, the body changes shape and the fix is different:
{ "error": { "code": 400, "message": "API key not valid. Please pass a valid API key.", "status": "INVALID_ARGUMENT", "details": [ { "@type": "type.googleapis.com/google.rpc.ErrorInfo", "reason": "API_KEY_INVALID" } ] } }
A `200` means the key works and the project is billed. That is the only signal worth trusting here, and it takes one API call to get.
Once you are on a paid project, four published levers reduce what a workload costs. None of them is free, and three of them are worth more than a discount code.
The introductory rates are the fifth lever and they expire. Google's pricing page lists $2 per million input and $10 per million output tokens through November 30, 2026, then $4 and $20 from December 1, 2026, with cache pricing doubling from $0.10 to $0.20 at the same time.
Here is what that expiry means in arithmetic, using only the published rates: $300 of credit buys 150 million input tokens at $2 per million, and 75 million at $4 per million. That is a calculation from two published numbers, not a quote, and it ignores output tokens, which cost five times more per token and will dominate any real workload.
The structural version of the same idea is to stop using a frontier model for everything. Argon is the wrong tool for classification and extraction at $2 per million input tokens when a small model costs $0.10 for the same work, which is the argument in the Haiku 5.5 subagent guide and in model and skill routing.
Rate limits are applied per project rather than per API key, and Google measures three dimensions: requests per minute, tokens per minute on input, and requests per day. Daily quotas reset at midnight Pacific time, and exceeding any single dimension triggers a rate limit error even when the others are unused.
Google's rate limits page lists tier structure and batch enqueued token ceilings, including 5,000,000 for Gemini 4 Argon on Tier 1 and 500,000,000 on Tier 2. The per-model requests per minute values are shown per project in AI Studio rather than published as a single table, so the honest way to know your ceiling is to read it in the console for your project.
The documented responses to expect are `429 RESOURCE_EXHAUSTED` and `503 UNAVAILABLE`, and Google's guidance is exponential backoff. The official SDKs already do this: the Python SDK retries transient errors up to four times, starting at roughly one second and capping at 60 seconds. If you are calling REST directly, you own that logic.
Two limits are specific to Argon and catch people who assume parity with other Gemini models. Explicit context caching is not supported, only implicit caching, so a pipeline built around a pre-warmed explicit cache will run and cost more without printing anything. And the model is not on the free tier at all, so there is no zero-cost lane to develop against: the development loop and the production loop are the same billed project.
If the goal is to evaluate Argon without a budget line, the $300 new-customer credit is the answer, and it is worth spending that credit where the model is distinctively strong rather than on chat. Long generation over an existing codebase is that case: the 200,000 line Python refactor Google published is the shape of task where the unified token budget matters and a small model cannot substitute.
If the goal is to run agents every day, the cheap model is the answer for most calls. The comparison with Claude Code has the price table side by side, the opencode setup shows how to keep the model choice reversible, and what Gemini 4 Argon is covers the lineup. For the tooling around it, the Jev opencode guide and the complete Jev guide cover the harness side, and why AI projects fail in small business covers the part where a pilot never reaches production.
If you want this costed for your own workload before you commit to a tier, that is the work we do: we embed AI engineers inside your existing engineering team, spec first, with the review gates and the exit written down. Send a written intake or see what we do.
Is there any free way to call Gemini 4 Argon?
Only through the $300 Google Cloud credit for new customers. The Gemini API free tier does not include Argon, and Google states that it is not available on the Google AI Pro plan either.
Does a Google AI Pro subscription give me API access to Argon?
No. The subscription includes Argon inside the Gemini app in selected countries. The API pricing page lists Argon as unavailable on that plan, so API calls need a billed Cloud project.
Which plan do I need for Antigravity?
Google AI Ultra. The launch post names Ultra subscribers as the Antigravity audience, and the Enterprise path runs through eligible Gemini Enterprise subscriptions.
What happens when the introductory price ends?
Input goes from $2 to $4 per million tokens and output from $10 to $20 on December 1, 2026, with cache reads moving from $0.10 to $0.20. Anything you costed before that date should be costed again after it.
Does Google train on my paid traffic?
The pricing page lists "Used to improve our products" as No on the paid tier. Read that row alongside your own compliance requirements rather than treating it as legal advice.
Can I cache a big system prompt to save money?
Only implicitly. Argon supports implicit caching for shared prefixes of 6,144 tokens or more and does not support explicit context caching, so structure repeated work around identical prefixes instead of a cache object.
[1] Google, Gemini Developer API pricing, read October 15, 2026. The statement that Argon is not available on the free tier or the Google AI Pro plan, the introductory and December rates, batch pricing, cache pricing, the 5,000 free grounding requests per month, and the data-use row.
[2] Google, Rate limits, read October 15, 2026. The per-project application of limits, the RPM, TPM and RPD dimensions, the midnight Pacific reset, tier upgrade behaviour, batch enqueued token ceilings for Argon, and the documented 429 and 503 responses.
[3] Google, Gemini API quickstart, read October 15, 2026. Key creation in AI Studio, setting `GEMINI_API_KEY`, and the billing step required to reach the paid tier.
[4] Google Cloud, Free cloud features and trial, read October 15, 2026. The $300 credit for new customers and the point at which charging begins.
[5] Google, Google AI plans and pricing, read October 15, 2026. The plan lineup, which plans list Gemini 4 Argon among their benefits, and the access multipliers on Pro and Ultra.
[6] Google, Introducing Gemini 4 Argon, October 13, 2026. Availability by surface and the introductory pricing.
[7] Google, Gemini 4 Argon comes to the Gemini app, October 13, 2026. Subscriber access in the app and AI Mode in Search.
[8] Google, Gemini 4 Argon model guide, read October 15, 2026. The enterprise client form, regions, thinking levels, implicit caching detail, the 95% implicit cache discount, the SDK versions, and the continuation-token behaviour behind `initial_prompt_token_count`.
[9] Google, Optimizing a 200,000-line Python codebase overnight with Gemini 4 Argon, October 13, 2026. The long-generation use case referenced in the closing section.
Maxpertise is an AI-native engineering company. We embed native AI engineers inside your team, live in about 10 days. Please enable JavaScript to view the site, or email contact@maxpertise.net.