GoogleGoogle·💬 Text Generation·New

Gemini 4 Carbon

ReasoningVisionFunction CallingWeb SearchAudioanonymized
🧠 Try in Intelligence →Try on Venice.ai ↗
Quick reference
Gemini 4 Carbon — TLDR
  • 🏢 Google DeepMind frontier model, released November 10, 2026.
  • 🧠 Built on Gemini 4 Argon, tuned for long-horizon agentic workflows.
  • 📏 Unified 1,048,576-token limit shared across input, thinking, and output.
  • 🔧 Tunable thinking levels low, medium (default), and high; long-decode continuation.
  • 👁️ Accepts text, images, audio and video; function calling and built-in tools.
  • 🔒 Defensive cybersecurity variant gated to verified defenders via the Fairwind Program.
  • ⚡ Spends more reasoning tokens than Argon for deeper verification loops.
💰 Pricing
$2.50 / $12.50
per 1M · input / output
📏 Context
1M tokens
📅 On Venice since
Nov 10, 2026
6 days ago
Provider

Google is an American multinational technology corporation and one of the world's most valuable brands. A subsidiary of parent company Alphabet Inc., Google operates across search, cloud computing, consumer electronics, and artificial intelligence. Its…

Read full profile →
41 models on Venice
15 text · 15 video · 4 image · 4 inpaint · 1 music · 1 embedding · 1 tts
Since Oct 15, 2024

About this model

Gemini 4 Carbon is Google's frontier text model aimed at sustained, multi-step agentic work across software engineering, enterprise knowledge work, scientific reasoning and cybersecurity defense. According to its model card, Carbon is based directly on Gemini 4 Argon, sharing that model's architecture and training data, with Carbon positioned as the heavier-reasoning member of the Gemini 4 family.

The clearest generational change versus Argon is how much deliberation the model applies: Google describes Carbon as using more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops, making Argon the recommended choice for faster interactive coding and everyday feature work while Carbon handles architecture, coordination and final verification. Both expose a unified 1,048,576-token budget that counts input, thinking traces, function calls and final output together — a departure from the separate input and output limits of earlier releases such as Gemini 3.1 Pro Preview — plus long-decode continuation for very large outputs. Thinking is controlled by discrete levels (low, medium, high) rather than a numeric budget, and sampling parameters are ignored.

Multimodal input spans images, audio and video alongside text, with function calling and built-in tools such as web search available through the API. A separate cyber-permissive configuration, defaulting to high thinking, is restricted to verified defenders under Google's Fairwind Program. Within the wider lineup, lighter or latency-sensitive workloads can be routed to Gemini 3.8 Flash.

This About section is AI-generated from public sources (Claude Opus 5), with no human editing. It may contain inaccuracies — verify critical details against the sources listed above.

Data sources: Venice API · HuggingFace · Wikipedia — enrichment updated 3d ago