Gemini 4 Carbon is now available
Gemini 4 Carbon, Google’s newest frontier LLM, delivers top‑tier performance on real‑world benchmarks while costing less than half of competing models. It ranks #1 on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench with >2× lower cost per task, and #2 on the new Multimodal Knowledge Agent benchmark with >6× lower average cost. Engineers can tap the model via the Google AI Studio API at $2 / M input tokens and $10 / M output tokens and leverage the /swarm feature in Google Antigravity to orchestrate multi‑agent workflows for complex coding, scientific research, and multimodal document processing.
⚡ Key Takeaways
- #1 ranking on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench with >2× lower cost per task versus the next best external model.
- Introduces the /swarm capability in Google Antigravity, enabling a single prompt to spawn a coordinated team of Carbon agents.
- Pricing: $2 per million input tokens and $10 per million output tokens, identical to Gemini 4 Argon.
- Multimodal Knowledge Agent (MKA) benchmark placement: #2 with >6× lower average cost per task than the top model.
- Demonstrated use: a multi‑agent Carbon team generated a 906 K‑triangle Three.js reconstruction from seven PDFs (206 engineering drawing views) via /swarm.
The combination of frontier reasoning, multimodal support, and sub‑linear cost lets production teams build sophisticated, agent‑driven pipelines—such as automated code debugging or large‑scale document analysis—without prohibitive inference spend.
✅ Practical Steps
- Call the Gemini 4 Carbon model through the Google AI Studio API using the standard pricing tier ($2 / M input, $10 / M output).
- Enable Google Antigravity’s /swarm endpoint to create a coordinated team of Carbon agents for multi‑step tasks.
- Assign Carbon agents to high‑reasoning roles (e.g., architectural design, verification) while pairing them with lighter models in parallel workspaces for efficiency.
Want the full story? Read the original article.
Read on Google DeepMind ↗