Gemini 4 Carbon
Overview
Gemini 4 Carbon is a Google DeepMind reasoning model in the Gemini 4 family, publicly released on November 10, 2026. Built on Gemini 4 Argon, it emphasizes deeper reasoning and extended verification for difficult software engineering, scientific research, and multimodal knowledge work.
Key Features
- Sustained Engineering Workflows: Handles complex debugging, architectural changes, and migrations through repeated investigation, implementation, and verification.
- Scientific Reasoning: Supports mathematical proofs, experimental data analysis, hypothesis exploration, and simulation development.
- Multimodal Analysis: Combines document text, charts, images, audio, and video to reason across varied source material.
- Configurable Thinking: Offers three reasoning levels to balance depth, response time, and token consumption.
- Extended Generation: Supports long decode for problems that need more than the default 64K-token generation budget.
Best Use Cases
- Complex Repository Maintenance: Investigating bugs across subsystems, planning migrations, and checking changes against existing tests.
- Document-Based Analysis: Synthesizing dense reports, tables, charts, and technical drawings into structured findings and report drafts.
- Research and Simulation: Exploring mathematical or scientific problems, analyzing experimental results, and developing computational models.
- Full-Stack and Visual Development: Building applications, refining interfaces, and implementing interactive 3D scenes with tool-assisted verification.
Capabilities and Limitations
| Capability | Description |
|---|---|
| Reasoning | Supports low, medium (default), and high thinking levels. |
| Multimodal | Accepts text, images, video, audio, and PDFs; produces text. |
| Context Window | 1,048,576 tokens shared across input, thinking, and output. |
| Generation Limit | 65,536 combined thinking and output tokens by default; long decode can use the remaining unified context budget. |
| Tool Use | Supports function calling, code execution, File Search, URL context, Search and Maps grounding, and computer use. |
| Structured Outputs | Supports schema-constrained structured responses. |
| Video Understanding | Supports agentic video understanding; video processing defaults to static. |
| Prompt Caching | Google's API enables implicit caching by default for prompts of at least 4,096 tokens; explicit caching is unsupported. |
| Knowledge Cutoff | March 2026; knowledge in some domains may be less current. |
Known Limitations
- Reasoning Budget and Latency: Carbon generally spends more thinking tokens per turn than Argon. A small generation budget can leave too little room for the answer, and difficult prompts may take several minutes or time out.
- Long Decode Requirements: Generation beyond 65,536 tokens requires a higher output limit and background execution or client-side continuation. Exact output-length targets are unreliable.
- Unsupported Controls and Outputs:
temperature,top_p, andtop_kare ignored. Native image generation, native audio generation, and the Live API are unsupported. - Cybersecurity Variants: Specialized cyber capabilities and Google's reported cyber benchmark results apply to verified-defender access through the Fairwind Program.
- Reliability: Hallucinations remain possible; factual claims, scientific conclusions, and generated code still require verification.
Standard Pricing
The token prices below are shown in USD per 1 million tokens; web search is billed in USD per use.
| Model | Input (USD / 1M Tokens) | Cache Write (USD / 1M Tokens) | Cache Read (USD / 1M Tokens) | Output (USD / 1M Tokens) | Web Search (USD / use) |
|---|---|---|---|---|---|
| Gemini 4 Carbon | $4.00 | $4.00 | $0.20 | $20.00 | $0.014 |
Credits settlement: B.AI converts charges at 1 USD = 1,000,000 Credits and deducts Credits from the account balance.
Web search costs 14,000 Credits per use, equivalent to $0.014 per use.
For this model on B.AI, cache writes are priced at 1x the input rate. This model supports implicit caching only; the cache-write price shown in the table does not imply support for explicit cache creation. Google's implicit caching requires prompts of at least 4,096 tokens; meeting that threshold does not guarantee a cache hit on every request.
Model prices in the documentation are based on the model providers' official list prices and serve as B.AI standard reference prices. B.AI may offer discounts or limited-time promotions on this basis; offer details are shown in Promotions and Pricing Updates and the website's discount groups. Applicable prices, offer terms, and final charges are subject to the platform display and actual billing records.