Gemini 4 Argon

​

Model specifications

​

Quick examples

​

Gemini native API

curl "https://api.mixroute.ai/v1beta/models/gemini-4-argon:generateContent" \
  -H "Authorization: Bearer $MIXROUTE_API_KEY" \
  -H "Content-Type: application/json" \
  --data '{
  "contents": [
    {
      "role": "user",
      "parts": [
        {
          "text": "Explain how to safely retry a failed database transaction."
        }
      ]
    }
  ],
  "generationConfig": {
    "maxOutputTokens": 32768,
    "thinkingConfig": {
      "thinkingLevel": "medium"
    }
  }
}'

​

OpenAI-compatible Chat Completions

curl "https://api.mixroute.ai/v1/chat/completions" \
  -H "Authorization: Bearer $MIXROUTE_API_KEY" \
  -H "Content-Type: application/json" \
  --data '{
  "model": "gemini-4-argon",
  "messages": [
    {
      "role": "user",
      "content": "Explain how to safely retry a failed database transaction."
    }
  ],
  "max_tokens": 32768
}'

​

Request parameters

​

Gemini native API

​

OpenAI-compatible Chat Completions

​

Integration notes

• The GenerateContent limits distinguish direct non-streaming REST calls (32,768 tokens per request) from streaming calls (65,536). These are provider limits; MixRoute route limits also apply.
• The 1M context window is not a 1M single-response allowance. Longer generation requires a continuation workflow; Google’s OpenAI compatibility endpoint does not support long decode. This page covers standard MixRoute requests, not Interactions or continuation support.
• Use thinking levels instead of the deprecated thinkingBudget. Do not combine both. Native thinkingLevel does not accept minimal, and thinking cannot be disabled.
• Omit temperature, topP, and topK (top_p and top_k in snake_case): Google ignores these sampling parameters for Argon. Explicit context caching via cachedContent is not supported.
• A small output budget can be consumed entirely by thinking, leaving a truncated or empty answer. Lower the thinking level to reduce reasoning effort rather than treating the token limit as an answer-length target.