Gemini 4 Carbon: Access, Price and Benchmarks
Gemini 4 Carbon is a Google model for cyber defense and long tasks, announced 28 Oct 2026; access is limited to cyber defenders, no developer date.
Google calls it its strongest model. We read the announcement and Google's other Gemini pages on 2 November 2026. Every benchmark figure below is Google's; we have not run these benchmarks ourselves.
How we chose
This is a launch report, not a ranking. We included only what Google's own pages state, and we label every benchmark "according to Google." What weighed most: access, pricing structure, and what a reader can do today. What did not weigh: any score or ordering of our own. What we did not measure: model quality, speed or cost on real workloads; we ran no tests. Our general rules are on the methodology page.
What is Gemini 4 Carbon?
Google describes Gemini 4 Carbon as "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." It builds on Gemini 4 Argon, which Google announced on 30 September 2026 and made available to developers on 13 October 2026, and Google says it carries the same multi-layered safety and security stack as Argon (training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring).
Google positions Carbon for "much harder, long-horizon engineering problems" and says it is "built for sustained agentic workflows." It also says internal teams saw it run for days on multi-step tasks. For the wider model family, see our entry, the radar of upcoming launches, and the latest news.
Gemini 4 Carbon release date and who can use it
According to Google, access starts the week of 28 October 2026 for "a subset of cyber defenders," and will "gradually phase access to additional trusted partners through our Fairwind Program." Carbon is not yet available to developers, businesses or consumers.
Google gives no release date for wider availability. It says it will iterate on guardrails first and then make Carbon available to developers, businesses and consumers through its product surfaces "as soon as possible." Google also says it is engaged in the U.S. government's voluntary process for pre-release model testing.
Fairwind access is requested through a Google Cloud account team, according to the Fairwind page, which said on 2 November 2026 that it works with over 650 partners globally. That page names Argon and does not mention Carbon, so the link between Carbon and Fairwind rests on the announcement post alone.
Gemini 4 Carbon pricing
Google's announcement sets an introductory price per million input and output tokens, and says cached input tokens are priced at 95% off the input price. After the introductory period, the announcement's footnote says the rates double for both input and output. We leave the dollar figures to Google's pages, because the introductory rates are time-limited: see the Gemini API pricing page for the current schedule, and the Carbon announcement for Carbon's numbers.
The Carbon post gives no end date for the introductory period. Google's API pricing page states an end date for Argon only (30 November 2026); we do not infer one for Carbon. The introductory rate is the same one Google gave Argon. Because Carbon is not yet sold to developers, these are announced rates, not a price list you can order from today.
Gemini 4 Carbon benchmarks (according to Google)
All figures below are reported by Google in its announcement. The post gives no methodology or links for its internal benchmarks, and the cyber results are described as evaluated without guardrails. Rankings are as of Google's 28 Oct 2026 post.
| Benchmark | What it measures | Result reported by Google | Source |
|---|---|---|---|
| Artificial Analysis Intelligence Index v5.0 | General intelligence index (third-party, result reported by Google) | #1, score 59 (index points), as of Google's 28 Oct 2026 post | Google announcement |
| SEC-bench Pro | Finding vulnerabilities | 88.2% | Google announcement |
| ExploitGym | Exploitation | 54.8% | Google announcement |
| CWE-bench | Patching | 71.0% | Google announcement |
| WebDev Arena | Web development (Elo) | #4, 1768 Elo, as of Google's 28 Oct 2026 post; +89 Elo (baseline not stated by Google) | Google announcement |
| Gray Swan IPI | Prompt-injection attack success (lower is better) | 0.0% (Argon: 0.7%) | Google announcement |
| Terminal-Bench 4.0 | Cheating by the model (lower is better) | 0.0% successful cheating, 0.4% cheat attempts | Google announcement |
Google also reports that its Chrome team found Carbon "more than 10X better than another lab" on high-quality patch generation. The comparison model is not named, so treat it as an unverifiable claim.
Gemini 4 Carbon vs Gemini 4 Argon
The table shows what each Google source says, nothing more. "Not stated" means the pages we read do not say; it is not a claim that the thing does not exist.
| Gemini 4 Carbon | Gemini 4 Argon | Source | |
|---|---|---|---|
| Announced | 28 Oct 2026 | 30 Sep 2026 | Google announcements |
| Access today | Subset of cyber defenders, phased via Fairwind | Trusted cyber defenders via Fairwind; paid tier of the Gemini API since 13 Oct 2026 | Google announcements, Argon developer post |
| Introductory price per 1M tokens | Same as Argon, per the Carbon post | Set in the Argon post and on the API pricing page | Google announcements |
| Price after introductory period | Double the introductory rate; end date not stated | Double the introductory rate from 1 Dec 2026 | Carbon post, Gemini API pricing page |
| API model code | Not listed in the Gemini API docs we fetched | gemini-4-argon | Gemini API models page |
| Context window | Not stated | 1,048,576 tokens (1M) | Argon model page |
| Pitch | Hard long-horizon engineering, cyber defense | Software engineering, enterprise knowledge work (legal, finance), cyber defense | Google announcements |
Google's own Carbon-over-Argon comparisons, all from its internal testing:
- More than 2x uplift in the success rate of generating secure code.
- Unsupported claims cut by approximately 2x.
Google also reports that Carbon "simultaneously reduced overreach by 26% and under-service by 29%." The sentence does not state what these are measured against, and we have not inferred it.
On CWE-bench, Argon's post cites 68% on CWE-bench v1 and the Carbon post cites 71.0% on Collinear's CWE-bench without a version. Do not read that as a like-for-like gain.
Which should you use today?
Our reading of the sources: if you need a Gemini 4 model through the API today, Argon is the one Google documents for developers, with a model code and a context window. Carbon has neither in the documentation we fetched, and Google says it is restricted to a subset of cyber defenders. Google offers no official guidance on choosing between the two.
Our reading: unless you are in the Fairwind Program, Argon is the only one of the two you can plan around now, and the practical step for Carbon is to watch for a developer announcement.
What Google hasn't said yet
- No API model code, context window, maximum output, supported modalities or rate limits for Carbon. It does not appear in the Gemini API models list or pricing page we fetched on 2 November 2026.
- No Vertex AI listing or regional availability for Carbon.
- No model card for Carbon. The Gemini 4 safety report that the post links is described there as Argon's report, and we did not read it.
- No date, or plan, for the Gemini app or consumer plans beyond "as soon as possible."
- No end date for Carbon's introductory price.
- No methodology or links for the internal benchmark comparisons with Argon.
- No stated position on which model to choose.
FAQ
What is Gemini 4 Carbon?
It is Google's model for cybersecurity and long-horizon engineering, announced on 28 October 2026, which Google calls its strongest.
When can I use Gemini 4 Carbon?
Only a subset of cyber defenders have access so far, widening through the Fairwind Program. Google gives no date for developers, businesses or consumers.
How much does Gemini 4 Carbon cost?
Google sets an introductory rate per 1M tokens that doubles afterwards, with cached input at 95% off the input price; the numbers are in the Carbon announcement.
Gemini 4 Carbon vs Gemini 4 Argon: which should I use?
As of 2 November 2026, Argon is the one with developer access and API documentation; Carbon is restricted. Google reports Carbon ahead of Argon on several internal tests.
How reliable are the Gemini 4 Carbon benchmarks?
They are Google's reported figures, and Google publishes no methodology for the internal ones. Check them against independent results once available.
Sources
- Introducing Gemini 4 Carbon (Google)
- Gemini 4 Argon (Google)
- Build with Gemini 4 Argon (Google)
- Gemini API models (Google AI for Developers)
- Gemini 4 Argon model page (Google AI for Developers)
- Gemini API pricing (Google AI for Developers)
- Fairwind Program (Google DeepMind)
Last reviewed: 2026-11-02.
Update trigger: a Carbon developer or Vertex AI announcement, a Carbon entry in the Gemini API docs, or a change to Google's pricing page.