Gemini 4 Carbon is Google's newest Gemini 4 model, released in November 2026 with different access rules on each Google surface and a price that doubles on 1 December 2026, so teams need exact facts before they build on it. This profile covers Gemini 4 Carbon (gemini-4-carbon) and its allowlisted cybersecurity variant, Gemini 4 Carbon for cyber (gemini-4-carbon-cyber-permissive). It sets out the model ID, release date, context window and output limits, prices before and after the introductory period, supported tools, regions, access restrictions, Google-reported benchmarks, the model card and data controls. It is written for developers, architects, security teams and buyers comparing Gemini 4 models.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Google's public documentation as of November 2026. Model details change, so confirm against Google's current documentation.
Key takeaways
- Released on 10 November 2026. Google Cloud lists Gemini 4 Carbon as GA with a release date of 10 November 2026, the same day as Google's "Gemini 4 Carbon is now available" announcement.
- Access differs by surface. Paying developers can use it through the Gemini API in Google AI Studio; on Google Cloud it is for verified enterprise customers in select regions; Google AI Ultra subscribers get it in the Gemini app, Antigravity and AI Mode.
- Introductory price is $2 input and $10 output per million tokens. Google's announcement states this rises to $4 and $20 after the introductory period ends on 30 November 2026.
- The context window is 1,048,576 tokens, shared with output. A single segment returns up to 65,536 tokens streamed or 32,768 non-streamed; long decodes can extend output on the v1beta1 API.
- A cyber variant is allowlist only. gemini-4-carbon-cyber-permissive is for defensive security work through Google's Fairwind Program, in the global and us regions.
What is Gemini 4 Carbon?
Gemini 4 Carbon is a proprietary multimodal large language model from Google, released as generally available on 10 November 2026. Google Cloud's model page describes it as Google's frontier model for deep work on complex, real-world tasks in coding, finance, legal, cybersecurity and science, and Google's announcement calls it "our strongest model". It sits in the Gemini 4 line with Gemini 4 Argon and Gemini 4 Hydrogen.
The Gemini 4 Carbon model card states that Carbon is based on Gemini 4 Argon and refers readers to the Argon model card for architecture, training data and hardware details. Google's announcement positions Argon as the better fit for most software engineering and knowledge work and Carbon for harder tasks such as complex multi-step coding and debugging and open-ended scientific research, noting that Carbon uses more tokens than Argon to explore and verify more thoroughly. For the full Google line-up see Google Gemini Models, and for the sibling model see the Gemini 4 Argon profile.
On Google Cloud, Carbon is designated a Google Frontier Model subject to Section 34 of the Service Specific Terms, and the cyber variant a Google Frontier Cyber Model subject to Section 31.
Key facts at a glance
As of 2026-11-15. Source: Google Cloud Gemini 4 Carbon model page, Gemini 4 Carbon model card and Google's announcement.
| Field | Value | Source |
|---|---|---|
| Provider | Google Cloud model page; model card; announcement | |
| Model ID (Gemini 4 Carbon) | gemini-4-carbon | Google Cloud model page; model card; announcement |
| Model ID (Gemini 4 Carbon for cyber) | gemini-4-carbon-cyber-permissive | Google Cloud model page; model card; announcement |
| Release date (Gemini 4 Carbon) | 2026-11-10 (GA) | Google Cloud model page; model card; announcement |
| Release date (Gemini 4 Carbon for cyber) | Not stated by Google on the pages checked (2026-11-15); offered to defenders about two weeks before Carbon GA, per the announcement | Google Cloud model page; model card; announcement |
| Status (Gemini 4 Carbon) | GA | Google Cloud model page; model card; announcement |
| Status (Gemini 4 Carbon for cyber) | Generally available behind an allowlist | Google Cloud model page; model card; announcement |
| Context window (Gemini 4 Carbon) | 1,048,576 tokens, unified input and output | Google Cloud model page; model card; announcement |
| Max output (Gemini 4 Carbon) | 65,536 per segment streamed, 32,768 non-streamed; up to 1,048,576 minus input with long decode (v1beta1) | Google Cloud model page; model card; announcement |
| Modalities (in / out) (Gemini 4 Carbon) | Text, image, audio, video / text | Google Cloud model page; model card; announcement |
| Thinking levels (Gemini 4 Carbon) | LOW, MEDIUM (default), HIGH | Google Cloud model page; model card; announcement |
| Thinking levels (Gemini 4 Carbon for cyber) | LOW, MEDIUM, HIGH (default) | Google Cloud model page; model card; announcement |
| Knowledge cutoff (Gemini 4 Carbon) | March 2026 | Google Cloud model page; model card; announcement |
| Based on (Gemini 4 Carbon) | Gemini 4 Argon (model card) | Google Cloud model page; model card; announcement |
| Based on (Gemini 4 Carbon for cyber) | Gemini 4 Carbon | Google Cloud model page; model card; announcement |
| Open weights / licence | No / proprietary | Google Cloud model page; model card; announcement |
| Parameters | Not stated by Google | Google Cloud model page; model card; announcement |
Google's comparison table for the cyber variant lists only its differences from Carbon (thinking default, computer use, regions, access), which is why other rows refer back to Carbon.
Release and version history
As of 2026-11-15. Source: Google Cloud model page and Google's announcements of 10 November 2026.
| Date | Event | Model IDs |
|---|---|---|
| About two weeks before 2026-11-10 | Carbon offered to cybersecurity defenders through the Fairwind Program, per the GA announcement | gemini-4-carbon-cyber-permissive |
| 2026-11-10 | Gemini 4 Carbon GA on Google Cloud; announcement "Gemini 4 Carbon is now available"; developer post on building with Carbon in the Gemini API | gemini-4-carbon |
| 2026-11-30 | End of introductory pricing, as announced | gemini-4-carbon |
| 2026-12-01 | Announced price of $4 input and $20 output per million tokens applies | gemini-4-carbon |
Google publishes no dated snapshot suffix for Carbon on the pages checked: the model ID is gemini-4-carbon. Gemini 4 Argon was released earlier, and the announcement describes Carbon as following it.
How much does Gemini 4 Carbon cost?
As of 2026-11-15. USD per million tokens. Source: Google's Gemini 4 Carbon announcement. The Google AI Studio pricing page checked on 2026-11-15 listed Gemini 4 Argon but no Carbon row, so only the announced figures are shown.
| Item | Price | Notes | Source |
|---|---|---|---|
| Input | Through 30 Nov 2026: $2; From 1 Dec 2026: $4 | None | Google's Gemini 4 Carbon announcement |
| Output | Through 30 Nov 2026: $10; From 1 Dec 2026: $20 | None | Google's Gemini 4 Carbon announcement |
| Context caching | Not stated by Google on the pages checked (2026-11-15) | None | Google's Gemini 4 Carbon announcement |
| Batch | Not stated by Google on the pages checked (2026-11-15) | None | Google's Gemini 4 Carbon announcement |
| Free tier | Paid tier required (developer post) | None | Google's Gemini 4 Carbon announcement |
| Cyber variant | Not stated by Google on the pages checked (2026-11-15) | None | Google's Gemini 4 Carbon announcement |
- Same introductory price as Argon: the announcement states Carbon's introductory price matches Argon's. The AI Studio pricing page lists Argon at $2 and $10 through 30 November 2026 and $4 and $20 from 1 December 2026.
- Token use: Google's announcement and model card both note that Carbon uses more tokens than Argon, and more at higher thinking levels, so per-task cost depends on the thinking level chosen.
- Consumption options on Google Cloud: Provisioned Throughput, batch inference, Standard PayGo and Priority PayGo are supported; fixed quota is not. Google Cloud prices are on its pricing page, which this profile did not check.
Context window and output limits
| Limit | Value | Notes | Source |
|---|---|---|---|
| Context window | 1,048,576 tokens, unified, includes output tokens | None | Google Cloud model page |
| Output per foreground segment | 65,536 streaming; 32,768 non-streaming | None | Google Cloud model page |
| Long decode | Up to 1,048,576 minus input tokens on v1beta1 | None | Google Cloud model page |
| Request size | GenerateContent 500 MiB compressed; Interactions 256 MiB compressed (synchronous or streaming), 9.75 MiB uncompressed (deferred) | None | Google Cloud model page |
| Media per prompt | Up to 3,000 images and documents, 10 videos, 1,000 audio files | None | Google Cloud model page |
| Video and audio length | About 45 minutes of video with audio or about 1 hour without; about 8.4 hours of audio | None | Google Cloud model page |
| Implicit caching minimum | 4,096 tokens | None | Google Cloud model page |
Modalities and tools
- Input and output: text, image, audio and video in; text out.
- Supported: thinking, system instructions, structured output, implicit context caching, grounding (Google Search, Google Maps and others), code execution, function calling, URL context, computer use and agentic video understanding.
- Not supported: RAG Engine, tuning, the Gemini Live API and explicit context caching. Setting thinking_budget or thinking_level "MINIMAL" returns HTTP 400.
- Fixed sampling: temperature 1.0, topP 0.95, topK 64 and candidateCount 1; custom values are ignored.
- APIs: GenerateContent, Interactions, Count Tokens and Chat Completions.
- Cyber variant: computer use is not supported on gemini-4-carbon-cyber-permissive.
Computer use on Carbon brings the browser and desktop agent risks covered in the Computer Use and Browser Agents Guide.
Where is Gemini 4 Carbon available?
As of 2026-11-15. Source: Google's announcement, developer post and Google Cloud model page.
| Channel | Details | Source |
|---|---|---|
| Gemini API (Google AI Studio, Android Studio) | All paying developers; paid tier of the API | Announcement; developer post; Google Cloud model page |
| Gemini Enterprise Agent Platform (Google Cloud) | Verified enterprise customers in select regions; global, us and eu (ML processing in us and eu) | Announcement; developer post; Google Cloud model page |
| Gemini for business (Gemini Enterprise) | Via the API or the model dropdown, plus Google Antigravity and CodeMender | Announcement; developer post; Google Cloud model page |
| Gemini app and AI Mode in Search | Rolling out to Google AI Ultra subscribers | Announcement; developer post; Google Cloud model page |
| Google Antigravity | Google AI Ultra subscribers; /swarm multi-agent feature in preview | Announcement; developer post; Google Cloud model page |
| Carbon for cyber | Allowlist through the Fairwind Program; global and us regions | Announcement; developer post; Google Cloud model page |
Google's announcement also opened applications for scientists to receive free six-month access to Argon and Carbon through Antigravity's highest usage tier, subject to eligibility and selection. For API key and project controls see the Google Gemini API and AI Studio Security Guide and the Gemini Enterprise Security Guide.
Access restrictions
| Model or surface | Who can use it (as stated by Google) | How to request |
|---|---|---|
| gemini-4-carbon, Gemini API | Paying developers | Paid tier of the Gemini API |
| gemini-4-carbon, Google Cloud | Verified enterprise customers in select regions | Not stated by Google on the pages checked (2026-11-15) |
| gemini-4-carbon, consumer apps | Google AI Ultra subscribers | Google AI Ultra plan |
| gemini-4-carbon-cyber-permissive | Allowlist only, for defensive cybersecurity such as vulnerability discovery, reverse engineering, exploit verification and patch generation | Through a Google contact or the Fairwind Program; access to earlier Gemini cyber models does not carry over |
The restricted variant and the general model share a name, so access reviews should check the exact model ID a project is allowed to call.
Benchmarks (provider-reported, dated)
All results below are reported by Google in its 10 November 2026 announcement. Google gave rankings and relative figures for most external evaluations rather than scores; no independent leaderboard results are included on this page.
| Evaluation | Result reported by Google | Comparator in the same source | Source and date |
|---|---|---|---|
| CursorBench 4.0, Perplexity WANDR, Zapier AutomationBench | Ranked first on each | More than 2x lower cost per task than the next external model | Google announcement, 10 Nov 2026 |
| Artificial Analysis Multimodal Knowledge Agent (MKA) | Ranked second | More than 6x lower average cost per task than the top model | Google announcement, 10 Nov 2026 |
| Balyasny General Financial Intelligence (pre-release) | Top of 23 frontier model configurations | Not named | Google announcement, 10 Nov 2026 |
| Project Fortify (internal) | 100% recall on scanning, 95% precision on triage | Over 25% relative recall gain over Argon on open-source repositories | Google announcement, 10 Nov 2026 |
| Simulated cyberattack testing (safety) | 99% of attack trajectories blocked; per-attempt breach rate 0.8% | Argon: 3.1% | Google announcement, 10 Nov 2026 |
| Gray Swan automated cyberattack queries (safety) | 2.17% success | Argon: 9.86% | Google announcement, 10 Nov 2026 |
| Agentic honesty benchmark, 375 tasks (safety) | 92% passed | Argon: 70% | Google announcement, 10 Nov 2026 |
| Gray Swan IPI benchmark (prompt injection) | 0% success rate | Not stated | Google announcement, 10 Nov 2026 |
Google described Carbon as leading on SEC-bench Pro, ExploitGym and CWE-bench v1 without giving numeric scores, so those are not tabulated. The announcement itself notes that prompt injection robustness can still improve.
Model card and system card
Google publishes a Gemini 4 Carbon model card (November 2026) and a separate Gemini 4 Carbon Safety and Security Report, both linked from the announcement. The model card states:
- Basis: built on Gemini 4 Argon; architecture, training data and hardware are in the Argon model card.
- Inputs and outputs: text, images, audio and video in; text out; combined input-output window of up to 1M tokens; knowledge cutoff March 2026, possibly less current in some domains.
- Intended uses: complex multi-step coding and debugging, open-ended scientific research, high-precision knowledge work and data analysis, multi-agent engineering tasks, and cybersecurity defence, vulnerability discovery and patching.
- Known limitations: hallucinations, jailbreak resistance still being improved, occasional slowness or timeouts, and higher token use at higher thinking levels.
- Internal safety evaluations versus Argon: text-to-text safety +3.50pp, multilingual safety +3.55pp, image-to-text safety -1.00pp, tone +0.10pp, unjustified refusals +0.10pp.
- Frontier Safety Framework: the model card refers to the separate safety report and mentions strengthened safeguards for CBRN and cyber misuse; it does not state critical capability level outcomes itself.
See the Frontier AI Model Risk Frameworks Guide for how such reports are structured.
Data handling
- Google Cloud controls: data residency, customer-managed encryption keys (CMEK), VPC Service Controls (VPC-SC) and Access Transparency (AXT) are listed for online prediction, batch inference and context caching.
- Processing location: Google Cloud lists ML processing for Carbon in the us and eu multi-regions.
- Terms: Carbon is subject to Section 34 of the Google Cloud Service Specific Terms, and the cyber variant to Section 31.
- Gemini API retention and training use: Not stated on the Carbon pages checked (2026-11-15); confirm in the Gemini API terms for the paid tier before sending sensitive data.
Deprecation status
| Model or item | Details | Source |
|---|---|---|
| gemini-4-carbon | Status: GA; Discontinuation date: Not stated by Google on the pages checked (2026-11-15) | Google Cloud model page |
| gemini-4-carbon-cyber-permissive | Status: Generally available behind an allowlist; Discontinuation date: Not stated by Google on the pages checked (2026-11-15) | Google Cloud model page |
The scheduled change Google has announced is the price rise on 1 December 2026, not a retirement. Earlier Gemini cyber model access does not transfer to the Carbon cyber variant.
How Gemini 4 Carbon differs from other models in its family
As of 2026-11-15. Source: Google announcement, model card and AI Studio models page. Models are listed alphabetically.
| Model | Positioning as stated by Google | Context | Introductory price (in / out) |
|---|---|---|---|
| Gemini 4 Argon | "Flagship reasoning model" (AI Studio); better fit for most software engineering and knowledge work (Carbon announcement) | 1M shared, long decodes up to 1M | $2 / $10 to 30 Nov 2026 |
| Gemini 4 Carbon | "Our strongest model" (announcement); for harder multi-step coding, debugging and scientific research; uses more tokens than Argon | 1,048,576 unified | $2 / $10 to 30 Nov 2026 |
- Same window, different depth: both list a shared 1M window; the difference Google states is task difficulty and token use.
- Different Cloud access: Carbon on Google Cloud is for verified enterprise customers in select regions.
- Across providers: Anthropic also restricts a configuration of its frontier model for defensive cyber work; see the Claude Mythos (Preview, 5, 5.1, 5.2) profile and Frontier AI Models Compared.
Practitioner checklist
- Record gemini-4-carbon or the cyber variant ID explicitly in the AI inventory and in project allow lists.
- Budget for the 1 December 2026 price change and for higher token use at HIGH thinking level.
- Do not send thinking_budget or thinking_level MINIMAL; use LOW, MEDIUM or HIGH.
- Remove custom temperature and top-p settings from code, since Carbon ignores them.
- Choose the us or eu endpoint where processing location matters, and apply CMEK and VPC-SC on Google Cloud.
- Restrict which API keys and service accounts can call Carbon and review Gemini API key exposure. See the LLMjacking Guide.
- Apply to the Fairwind Program separately if defensive security work needs the cyber variant.
Standards and references
- Google Cloud: Gemini 4 Carbon model page (accessed 2026-11-15)
- Google DeepMind: Gemini 4 Carbon model card (accessed 2026-11-15)
- Google: Gemini 4 Carbon is now available (accessed 2026-11-15)
- Google: Gemini 4 Carbon for developers (accessed 2026-11-15)
- Google AI Studio: Gemini API pricing (accessed 2026-11-15)
- Google AI Studio: Gemini models (accessed 2026-11-15)
Related NHI Mgmt Group resources: Google Gemini Models, Gemini 4 Argon, Frontier AI Models Compared, Google Gemini API and AI Studio Security Guide.