Google's stable model ID is gemini-4-carbon. The Gemini API documentation provides an Interactions API call example. The model accepts text, image, video, audio, and PDF inputs, generates text, and uses a shared 1,048,576-token context window for input and output.
Suitable tasks: Configuring the model ID, input types, thinking level, and output limit according to the Google Gemini API documentation.
Unsuitable tasks: This information cannot establish whether a particular account, region, or third-party client is authorized to use the model.
Applicable model version: Stable Gemini 4 Carbon, gemini-4-carbon.
Applicable client, agent, or API: Gemini API; the official quickstart example uses the Interactions API.
Recommended reasoning level and parameters: The available thinking_level values are low, medium (default), and high. The default synchronous output limit is 65,536. For longer outputs, raise max_output_tokens and use background execution or foreground continuation.
The following is Google's official REST quickstart example. The original request structure and sample input are preserved:
curl "https://generativelanguage.googleapis.com/v1beta/interactions" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-H 'Content-Type: application/json' \
-X POST \
-d '{
"model": "gemini-4-carbon",
"input": "Write a three.js script that renders a realistic 3D black hole with gravitational lensing."
}'Set the request's model to gemini-4-carbon.
You can omit thinking configuration by default. When specifying it, Google's example uses generation_config.thinking_level, with values low, medium, or high.
max_output_tokens limits the combined token count for thinking and final output. To run long decode in the background, set background: true and raise the limit above 65,536. The server generates up to 64K tokens per segment. Combine the model_output text in interaction.steps.
For multi-turn conversations, use previous_interaction_id to retain server-side session state.
Model ID and version: The model page lists the Model code as gemini-4-carbon, and the Stable version is also gemini-4-carbon.
Modalities: Supported inputs are Text, Image, Video, Audio, and PDF; the output is Text.
Token limits: The shared input and output context is 1,048,576 tokens; the default synchronous maximum output is 65,536; long decode supports up to 1,048,576, subject to input token usage.
Capabilities: Supports agentic video understanding, code execution, computer use, context caching, file search, function calling, Google Maps/Search grounding, structured outputs, thinking, and URL context; does not support audio generation, image generation, or the Live API.
Caching limits: Explicit caching is not supported. Google's migration guidance says to remove cachedContents / caches.create and cached_content from requests.
Sampling parameters: temperature, top_p, and top_k do not apply to this model and are ignored if provided.
Reasoning and long-output parameters: The default thinking level is medium; thinking_level supports low, medium, and high. max_output_tokens counts thinking and output together; for long outputs, Google recommends using background: true. For complex reasoning or coding tasks, Google's documentation recommends starting at 131,072 or 262,144; the maximum is limited by the shared context window and input length.
Calling method: The official quickstart example uses POST https://generativelanguage.googleapis.com/v1beta/interactions, with model and input request fields.
Consumption options: Batch API, Flex inference, and Priority inference are all marked as supported.
Access and pricing: The official API documentation says Carbon is available on the Gemini API paid tier and is not available on the free tier. The introductory prices shown on the page are valid through 2026-11-30: $2 per million input tokens and $10 per million output tokens (including thinking tokens); after that, the listed prices are $4 / $20. Prices may change; refer to the current official pricing page.
Additional official API notes: What's new in Gemini 4 Carbon shows the thinking, long-output, caching, and migration parameters above; it is not evidence for another model or a third-party implementation.
The specifications come from Google's documentation and do not represent independent performance tests.
Account eligibility, available regions, quotas, current prices, and API behavior may change; this article records what was visible on the collection date.
The documentation recommends using background execution for long-running tasks, or setting the HTTP client timeout to at least 15 minutes.
Collected on 2026-11-16. Opened the user-provided ai.google.dev model page through Tabbit; that address redirects to the canonical Google AI Studio documentation URL. Read the model specification page and its official Carbon API notes, including the REST code example. No API requests or performance tests were conducted.
Gemini 4 Carbon