Gemini 4 Carbon
Google DeepMind's security-focused frontier model. Accepts text and an optional image plus a system prompt and reasoning effort, billed per token.
Gemini 4 Carbon is Google DeepMind's security-focused frontier model in the Gemini 4 series. Google reports it first on the Artificial Analysis Intelligence Index and state-of-the-art on vulnerability discovery, exploitation, and patching benchmarks, with stronger secure code generation than Gemini 4 Argon.Muapi exposes Gemini 4 Carbon through one simple endpoint with text and image input, an optional system prompt, a reasoning_effort control, token-based billing at $4.00 per million input tokens and $20.00 per million output tokens, and the same asynchronous submit-and-poll contract as every other model.
Gemini 4 Carbon is available now via API
Try Gemini 4 Carbon in the playground, or call it from your own code with a Muapi API key. See the Gemini 4 hub for Gemini 4 Argon and every other Gemini 4 model.
Google DeepMind's security-focused frontier model. Accepts text and an optional image plus a system prompt and reasoning effort, billed per token.
Gemini 4 Carbon is the security-focused model in Google DeepMind's Gemini 4 series. It ranks first on the Artificial Analysis Intelligence Index and, per Google, posts state-of-the-art results on vulnerability discovery (SEC-bench Pro), exploitation (ExploitGym), and patching (CWE-bench), while more than doubling Gemini 4 Argon's secure code generation success rate in Google's internal evaluation. Those figures are Google's own reported results, not Muapi measurements.
Muapi wraps it as a single gemini-4-carbon endpoint. Send a prompt with an optional image_url, system_prompt, and reasoning_effort (low, medium, or high), then poll the shared result endpoint. For the general-purpose Gemini 4 tier, see Gemini 4 Argon; for faster, cheaper everyday chat, see Gemini 3.8 Flash.
Built for finding, exploiting, and patching vulnerabilities, with strong results on security benchmarks as reported by Google.
Write and review code with fewer security flaws than the previous Gemini 4 tier, according to Google's internal evaluation.
Set reasoning_effort to low, medium (default), or high to trade speed and cost for depth on hard problems.
Include an image alongside your prompt to ask about screenshots, architecture diagrams, or scanned documents.
Pay $4.00 per million input tokens and $20.00 per million output tokens, with no subscription.
Audit functions and pull requests for injection, memory-safety, and authentication flaws, then draft patches.
Classify findings by CWE, explain exploitability, and prioritize remediation.
Plan and implement multi-step engineering work with strong reasoning.
Use Carbon as the reasoning core of multi-step agents and automated workflows.
Ask questions about screenshots, diagrams, and scanned pages by attaching an image.
| Feature | Gemini 4 Carbon | Gemini 4 Argon |
|---|---|---|
| Positioning | Security-focused frontier reasoning | General frontier reasoning |
| Context window | Not published by Google | 1M-token unified context and output |
| Input | Text, image, system prompt, reasoning effort | Text, image, system prompt |
| Price (input / output per 1M tokens) | $4.00 / $20.00 | $2.00 / $10.00 |
| Minimum charge per call | $0.0004 | $0.0002 |
| Best for | Vulnerability analysis and secure code | Long-horizon coding and deep analysis |
Submit a JSON request with your Muapi API key, save the request ID, and poll the shared result endpoint for the completed text.
curl -X POST https://api.muapi.ai/api/v1/gemini-4-carbon \
-H "x-api-key: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt":"Audit this function for injection vulnerabilities and propose a patch","system_prompt":"You are a senior application security engineer.","reasoning_effort":"high"}'
# Response: {"request_id":"REQUEST_ID"}Call gemini-4-carbon with a prompt, and optionally a system_prompt, image_url, and reasoning_effort.
curl https://api.muapi.ai/api/v1/predictions/REQUEST_ID/result \ -H "x-api-key: YOUR_API_KEY" # Poll until status is completed, then read the generated text from outputs.
Poll until the task is completed, then read the generated text from outputs.
Gemini 4 Carbon is Google DeepMind's security-focused frontier model in the Gemini 4 series. It ranks first on the Artificial Analysis Intelligence Index and is strong at finding, exploiting, and patching vulnerabilities as well as general coding and reasoning.
Billing is per token at $4.00 per million input tokens and $20.00 per million output tokens, with a minimum charge of $0.0004 per call. Reasoning tokens count as output. There is no subscription.
It sets how much the model reasons before answering: low, medium (default), or high. Higher effort improves depth on hard problems but is slower and uses more output tokens.
Yes. Pass an image_url with your prompt to ask questions about a screenshot, diagram, or photographed document.
The endpoint is asynchronous. You receive a request_id immediately, then poll the predictions result endpoint or provide a webhook URL to be notified when the text is ready.
Carbon is the newer, security-focused model with stronger vulnerability-finding and secure code generation results as reported by Google; Argon is the general frontier reasoning tier with a 1M-token context. Both use the same prompt, image_url, and system_prompt fields, and Carbon adds reasoning_effort.
No. The benchmark and evaluation figures cited on this page are Google's reported results. Run your own evaluations on your code before relying on the model for security decisions, and have a human review findings and patches.
Add security-focused frontier reasoning to your product with one API.