Gemini 4 Argon

Gemini 4 Argon (google/gemini-4-argon) is Google’s senior fourth-generation model. It accepts images and files, reasons with a controllable depth and has native web search. Thinking tokens are billed at their own rate — the same as output tokens.

Parameters and capabilities

Parameter	Value
model	google/gemini-4-argon
Images on input	yes (image_url, data URLs)
Files on input	yes (PDF/DOCX/TXT etc., up to 5)
Reasoning	reasoning_effort: low, medium, high; cannot be disabled
Thinking tokens	$12.5 per 1M — same rate as output
Native web search	yes — web_search: "native", $0.0175 per request
Cache read	$0.125 per 1M tokens
Cache write	$0.235 per 1M tokens
Sampling	temperature, top_p, seed

Example

Python

from openai import OpenAI

client = OpenAI(base_url="https://api.mixen.ai/v1", api_key="mxn-...")

resp = client.chat.completions.create(
    model="google/gemini-4-argon",
    messages=[{"role": "user", "content": "Compare two database architectures and recommend one"}],
    reasoning_effort="high",
)
print(resp.choices[0].message.content)

cURL

curl https://api.mixen.ai/v1/chat/completions \
  -H "Authorization: Bearer $MIXEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemini-4-argon",
    "messages": [{"role": "user", "content": "Compare two database architectures and recommend one"}],
    "reasoning_effort": "high"
  }'

Web search — via "web_search": "native"; sources arrive in message.annotations.

The Gemini wire protocol works too: base URL https://api.mixen.ai/v1beta and the x-goog-api-key header — the model is reachable from gemini-cli and the google-genai SDK, see the Gemini CLI guide.

Pricing

Per 1M tokens: input — $2.5, output — $12.5. Thinking — $12.5 per 1M. Cache: read — $0.125, write — $0.235. Showcase prices follow the live exchange rate — take fresh values from the catalog.

Strengths and limitations

• Native web search: sources in annotations, charged per request separately from tokens.
• Cache write is ten times cheaper than input ($0.235 vs $2.5) — a rare case where warming the cache pays back fast.
• Reasoning cannot be disabled: reasoning_effort: "off" is rejected; the minimum depth is low.
• Thinking is billed — keep headroom in max_tokens, or the answer gets truncated (finish_reason: "length").

All models — in the catalog. The general workflow — chat guide.