# Gemini 4 Argon API: key, first call and which tool to use

Gemini 4 Argon API how-to from Google's docs: get a paid key, make the first call with gemini-4-argon, the parameters that now fail, and grounding vs MCP.

## Summary

The Gemini 4 Argon API is the Gemini API with the model id `gemini-4-argon` on a billed project. Google made it generally available on 2026-10-13, and its docs show the call shape: the `google-genai` Python SDK, the Interactions API, and a key in `GEMINI_API_KEY`.

Old code breaks in small ways on Argon. Some sampling settings are ignored, a few now return HTTP 400, and the thinking setting has a new name. This post walks through the key, a first call, the parameters to remove, and how to choose between Google's search grounding, your own functions and an MCP server. By the end you have a working call and a plan for the tools your agent needs.

Last checked 2026-10-15, against Google's developer docs and looot's live catalog.

## How do you get a Gemini 4 Argon API key?

Create an API key in [Google AI Studio](https://aistudio.google.com/) and link a billing account to its project. Argon is not on the free tier, so a key on a project without billing cannot call it. Linking billing puts the project in Tier 1 on Google's [rate limits page](https://ai.google.dev/gemini-api/docs/rate-limits) (read 2026-10-15).

| Tier | How you reach it | Billing cap | Source |
|---|---|---|---|
| Tier 1 | link an active billing account | $250 | Gemini API rate limits, read 2026-10-15 |
| Tier 2 | $100 paid and 3 days since the first payment | $2,000 | Gemini API rate limits, read 2026-10-15 |
| Tier 3 | $1,000 paid and 30 days since the first payment | $20,000 to $100,000 or more | Gemini API rate limits, read 2026-10-15 |

The rate limits page does not print Argon's requests per minute. It sends you to AI Studio, where each project sees its own limits. What a call costs is in [Gemini 4 Argon pricing](https://looot.ai/blog/gemini-4-argon-pricing).

## Your first Gemini 4 Argon API call

Install the SDK, put the key in the environment, and send one short prompt. This is Google's [quickstart](https://ai.google.dev/gemini-api/docs/quickstart) (read 2026-10-15) with the model id changed from `gemini-3.8-flash` to `gemini-4-argon`.

The same call over REST goes to `https://generativelanguage.googleapis.com/v1beta/interactions` with the key in an `x-goog-api-key` header and a JSON body holding `model` and `input`. We have not run this exact call against Argon, so send one cheap prompt first and read the error if there is one.

The model takes text, images, video, audio and PDF as input and returns text. Its context window is 1,048,576 tokens shared by input and output, with 65,536 output tokens per request by default, per the [Gemini 4 Argon model page](https://ai.google.dev/gemini-api/docs/models/gemini-4-argon) (read 2026-10-15).

## Which parameters does the Gemini 4 Argon API reject?

Argon drops the old thinking budget and most sampling settings. Google Cloud's [Gemini 4 Argon developer guide](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/guides/gemini-4-argon) (read 2026-10-15) lists the rules below. It is written for Google Cloud's Agent Platform, so confirm each one on your own key.

| Setting | On Argon |
|---|---|
| `thinking_level` | `LOW`, `MEDIUM` (default) or `HIGH` |
| `thinking_budget`, or the level `MINIMAL` | HTTP 400 |
| `temperature`, `top_p`, `top_k` | ignored |
| `frequency_penalty`, `presence_penalty`, `candidate_count` | HTTP 400 |
| Explicit context caching | not supported, caching is implicit |
| A chat history that ends with a model turn | not supported, no prefilled answers |
| Function results | must match the id, name and order of each function call |

The guide also asks for `google-genai` 2.30.1 or newer. If a call that worked on an older Gemini model now fails with 400, remove the settings in this table first.

## Grounding, functions or MCP: which tool should Argon use?

Use Google's search grounding for answers backed by Google Search, your own functions for your own systems, and an MCP server for data neither one returns, such as company records or ranked Google positions. All three are listed as supported on the model page, and many agents combine them.

| Tool | What the model gets | What it costs | Source |
|---|---|---|---|
| Search grounding, `{"type": "google_search"}` | Google results with cited sources | 5,000 free a month, then $14 per 1,000 searches | Gemini API pricing, read 2026-10-15 |
| URL context | the text of a page you name | free as a tool, the text counts as input tokens | Gemini API pricing, read 2026-10-15 |
| Function calling | whatever your own code returns | your own cost | Gemini 4 Argon model page, read 2026-10-15 |
| A remote MCP server | the server's tools, over Streamable HTTP | the server's price per call | Gemini 4 Argon MCP |
| Code execution | Python run by Google | see Google's pricing page | Gemini 4 Argon model page, read 2026-10-15 |

Grounding is one line in the tools list, the same shape as Google's [grounding docs](https://ai.google.dev/gemini-api/docs/google-search) (read 2026-10-15):

Each search the model runs is billed, so read the `google_search_call` steps in the response to see how many it made. When grounding gets expensive at your volume, [Grounding with Google Search pricing](https://looot.ai/blog/grounding-with-google-search-pricing) works through the numbers.

## Add a looot data step to the Gemini 4 Argon API

looot is one MCP server that reaches many data providers, so Argon can look up a company, read Google results with positions or find a work email without a key for each vendor. The `mcp_server` tool goes in the same `tools` list as grounding. The full code is in [Gemini 4 Argon MCP](https://looot.ai/blog/gemini-4-argon-mcp).

Two jobs an Argon agent often needs are [company enrichment](https://looot.ai/use-cases/enrich-a-company), which takes a domain and returns the company's name, size and links, and [Google results with positions](https://looot.ai/use-cases/google-serp-api) for SEO work. The model reads each job's price before it runs it, and a miss at one provider moves to the next.

## What it cannot do

- looot does not resell Gemini 4 Argon. Argon is not in looot's catalog today; your key and your Google bill stay yours.
- It does not replace search grounding. Grounding answers from Google Search inside one call; looot adds data jobs grounding does not return.
- We have not run the calls in this post against Argon. They follow Google's quickstart and guide with the model id changed.
- It does not run on a timer. A recurring report needs scheduled runs, not available yet.
- The [Gemini 4 Argon vs GPT-6.1 Sol vs Claude Opus 5.5](https://looot.ai/blog/gemini-4-argon-vs-gpt-6-1-sol-vs-claude-opus-5-5) comparison covers the other two models; this post covers only Google's API.

## Questions

### What is the Gemini 4 Argon API model id?

`gemini-4-argon`, according to the [Gemini 4 Argon model page](https://ai.google.dev/gemini-api/docs/models/gemini-4-argon). Google Cloud also lists `gemini-4-argon-cyber-permissive` for approved security work.

### Is there a free Gemini 4 Argon API key?

No. Argon is not on the free tier, so the key's project needs a linked billing account. A free-tier key works for other Gemini models only.

### How do I set the thinking level in the Gemini 4 Argon API?

Set `thinking_level` to `LOW`, `MEDIUM` or `HIGH`; `MEDIUM` is the default. The old `thinking_budget` and the `MINIMAL` level return HTTP 400, per Google Cloud's developer guide.

### Why does my Gemini 4 Argon API request return 400?

The usual cause is a setting Argon no longer takes: `thinking_budget`, `MINIMAL`, `frequency_penalty`, `presence_penalty` or `candidate_count`. Remove them, and make sure the chat history does not end with a model turn.

### Does the Gemini 4 Argon API support MCP?

Google documents remote MCP servers as a tool type in the Interactions API, over Streamable HTTP, and Argon supports function calling. [Gemini 4 Argon MCP](/blog/gemini-4-argon-mcp) shows the setup with looot's server.

### AI Studio or Google Cloud for the Gemini 4 Argon API?

AI Studio gives you an API key and the Gemini API. Google Cloud's Agent Platform serves the same model id to Cloud projects with regional endpoints. The global price is the same on both.

### Should I call data providers directly instead of through looot?

You can, with one key, one schema and one bill per provider. One MCP server keeps the agent's setup to a single URL and lets a miss at one provider fall through to the next.

## For agents

### Code (bash)

```bash
pip install -U google-genai
export GEMINI_API_KEY=...   # a key on a project with billing linked
```

### Code (python)

```python
from google import genai

client = genai.Client()
interaction = client.interactions.create(
    model="gemini-4-argon",
    input="Explain how AI works in a few words",
)
print(interaction.output_text)
```

### Code (python)

```python
interaction = client.interactions.create(
    model="gemini-4-argon",
    input="What did Google announce for Gemini 4 Argon this week?",
    tools=[{"type": "google_search"}],
)
```

### Gemini 4 Argon API with a data step

```text
Use looot to give my Gemini 4 Argon API agent company and search data. First set up https://looot.ai/skill.md and add the MCP server https://api.looot.ai/mcp to the agent's tools. Then search the catalog for "enrich a company", inspect the cheapest job, and run job:company.enrich with fallback on for the domain blog.google. Run job:google.serp.organic for "gemini 4 argon api" and list the top five results with their positions. Tell me each price before the run and stop if one call would cost more than $0.05.
```