Gemini 4 Carbon by Google

Overview

Gemini 4 Carbon is Google DeepMind’s frontier Gemini 4 model for deep, multi-step work. It processes very long context, then produces text outputs like analyses, plans, code, and structured write-ups. You can feed it text and also attach images, audio, video, and office documents for grounded reasoning over real materials. It’s useful when you need one model to read lots of evidence, reason over it, and return a clean, ready-to-use result.

What you can build

• Long-form technical writing from mixed sources like PDFs, slides, and spreadsheets
• Multi-step debugging and refactors across large code snippets and design notes
• Security review support for code changes, with focus on defensive explanations and patch proposals
• Research synthesis that merges notes, screenshots, and meeting audio into a single brief
• Visual understanding workflows like chart and diagram interpretation with written conclusions
• Agent-style task plans that break big work into tracked steps and checklists

Inputs

• Your main instruction as plain text. Keep it specific about the deliverable format you want.
• Optional attachments for context, including: - Up to 10 images (each up to 7 MB).
  • Up to 10 videos (each up to 45 minutes).
  • An optional video sampling rate. Higher sampling uses more tokens but captures more motion detail.
  • Up to 1 audio file (up to 8.4 hours).
  • Documents such as PDF, plain text, Microsoft Office files (DOC, DOCX, PPT, PPTX, XLS, XLSX), and OpenDocument files (ODT, ODP, ODS), when file upload is enabled on the surface you’re using.
• Optional system-level guidance that sets global rules for the run, like tone, constraints, or output structure.
• Optional reasoning depth selection (low, medium, or high). Higher settings spend more tokens on deeper reasoning.
• Optional response length limit in tokens. This budget includes internal thinking and the visible answer.
• Optional web grounding toggles to let the model ground answers in Google Search and/or Google Maps results.
• Optional identifiers for user and session history, if you want the same conversation to persist across turns.
• Optional chat-style history as a JSON list of turns (role plus content). Use this instead of a single instruction when you want multi-turn behavior.
• Optional continuation inputs for very long tasks, where you resume from a prior run token and (if applicable) provide tool results for a continuation step.

Outputs

The model returns generated text.

That text often includes: - A direct answer to your prompt

• Structured sections (headings, bullet lists, tables, or step-by-step procedures) when you request a format
• Long-form outputs that can span tens of thousands of tokens in a single response segment

If the output hits your token limit, the response can end early. For long documents or long code generation, you may need to continue the output in follow-up turns.

Recommended settings

• Use medium reasoning depth for most coding, analysis, and document QA.
• Use high reasoning depth for hard debugging, deep math and science reasoning, or complex multi-file planning.
• Keep the response token budget high when you expect long outputs. Carbon shares a single token budget across input, thinking, and output.
• When sending video, raise the sampling rate only if motion detail matters. Lower rates reduce token use.

Limitations

• The model has a shared context window of 1,048,576 tokens across input, internal thinking, and output.
• A single foreground response segment typically caps at 65,536 output tokens for streaming and 32,768 tokens for non-streaming.
• The model can hallucinate. Treat outputs as drafts that still need verification.
• Knowledge cutoff is March 2026, so newer events and releases may be missing.
• Some sampling-style controls are ignored by Gemini 4 models, and some penalty-style controls can cause request errors on certain surfaces.
• Low-quality inputs reduce quality. Scanned PDFs, blurry photos, messy spreadsheets, and low-sample video can hide key details.
• Custom function tools may be restricted on some hosting surfaces, even if the underlying model family supports tool use.
• Availability can be limited by region and account type on some Google enterprise surfaces.

Safety & compliance

Gemini 4 Carbon includes safety and misuse safeguards, and some variants are gated for verified defensive security users.

Follow the applicable Google terms for the surface you use. Some enterprise deployments treat Gemini 4 Carbon as a “frontier model” with abuse monitoring requirements.

Don’t use the model to generate or automate harmful activity. For security work, keep outputs focused on defensive analysis and remediation.