Oleksandr Kuzmenko, EditorPublished 5 min readAI-assisted · editor-reviewed
TL;DR
- Gemini 4 Carbon is now available to all developers with support in the llm-gemini CLI plugin.
- Introductory rates ($2/M input, $10/M output) double on December 1st to $4/M and $20/M.
- Thinking effort defaults to medium, with low and high options available.
- Google Antigravity's /swarm feature showcases Carbon coordinating Blender and Three.js subagents.
Key facts
- Launch Input Price
- $2.00 / 1M tokens
- Launch Output Price
- $10.00 / 1M tokens
- Post-Dec 1 Input Price
- $4.00 / 1M tokens
- Post-Dec 1 Output Price
- $20.00 / 1M tokens
- Default Thinking Effort
- medium
Source: Key facts — Launch Input Price, Launch Output Price, Post-Dec 1 Input Price, Post-Dec 1 Output Price
Source: Key facts. Launch Input Price: $2.00 / 1M tokens. Launch Output Price: $10.00 / 1M tokens. Post-Dec 1 Input Price: $4.00 / 1M tokens. Post-Dec 1 Output Price: $20.00 / 1M tokens. The chart compares only values that share a unit. Each figure is copied from the story.
Metric chart
- $2.00 / 1M tokensLaunch Input Price
- $10.00 / 1M tokensLaunch Output Price
- $4.00 / 1M tokensPost-Dec 1 Input Price
- $20.00 / 1M tokensPost-Dec 1 Output Price
- Launch Input Price$2.00 / 1M tokens
- Launch Output Price$10.00 / 1M tokens
- Post-Dec 1 Input Price$4.00 / 1M tokens
- Post-Dec 1 Output Price$20.00 / 1M tokens
Public Access and Pricing Window
Previously restricted to cybersecurity specialists, Google has officially made Gemini 4 Carbon broadly available. Command-line users can access the model via Simon Willison's llm-gemini plugin.
Google launched Carbon with temporary aggressive pricing:
- Input tokens: $2.00 per million
- Output tokens: $10.00 per million
These rates match mid-tier frontier models like GPT-6.1 Sol and Claude Sonnet 5.5. However, this introductory window closes on December 1st, when rates jump to $4.00 per million input and $20.00 per million output (matching Claude Opus 5.5 pricing tiers).
Reasoning Modes and Multi-Agent Swarms
Gemini 4 Carbon introduces native thinking effort tiers (low, medium, and high), defaulting to medium. Early testing demonstrates strong spatial and frontend synthesis capabilities, generating standalone zero-dependency HTML and Three.js scenes from complex prompts.
In tandem, Google demonstrated orchestration workflows within the experimental Antigravity IDE. Using the /swarm feature, Carbon coordinates multiple specialized subagents: one models 3D geometries in Blender against Omni and Nano Banana references, while others compile Three.js browser logic, write WebGL shaders, and synthesize Web Audio sound effects.
When to use
- Complex interactive 3D and spatial coding synthesis.
- Workflows requiring subagent swarming across multiple creative tools.
When NOT to use
- For cost-sensitive applications post-December 1st where cheaper models suffice.
- Simple zero-reasoning tasks where thinking models add unnecessary latency.
What to do today
- Test Gemini 4 Carbon workloads prior to the December 1st price increase.
- Experiment with low, medium, and high thinking effort settings to balance token usage and accuracy.
- Review Antigravity /swarm capabilities for multi-agent asset and logic creation.
Step names are taken from What to do today.
Step names are taken from What to do today. Test Gemini 4 Carbon workloads prior to the December 1st price increase.. Experiment with low, medium, and high thinking effort settings to balance token usage and accuracy.. Review Antigravity /swarm capabilities for multi-agent asset and logic creation.
- Test Gemini 4 Carbon workloads prior to the December 1st price increase.
- Experiment with low, medium, and high thinking effort settings to balance token usage and accuracy.
- Review Antigravity /swarm capabilities for multi-agent asset and logic creation.