Gemini 4 Carbon
- Provider
- Status
- Restricted
- Context
- 1,000,000 tok
- Price
- $2 / $10 /MTok
Gemini 4 Carbon is Google’s frontier model, announced on 28 October 2026 and rolled out first to a small group of cyber defenders through Google’s Fairwind Program, with wider access promised and no date given (R&D World). Google describes it as its best model yet at finding and exploiting security vulnerabilities.
Artificial Analysis scored it 59 on the Intelligence Index v5.0 at high reasoning, the top score on that version, one point above Claude Opus 5.5 (58). Google announced an introductory price of $2/$10 per million tokens, doubling to $4/$20 after an introductory period it has not dated.
Quick specs
| Gemini 4 Carbon | |
|---|---|
| Announced | 28 October 2026 |
| Availability | Selected cyber defenders in Google’s Fairwind Program |
| Input price | $2.00 / MTok introductory, then $4.00 |
| Cached input | 95% off |
| Output price | $10.00 / MTok introductory, then $20.00 |
| Context window | 1,000,000 tokens |
| Input | Text, images, video, audio |
| Knowledge cutoff | Not published |
| AA Intelligence Index | 59 (v5.0, high) |
| Cost per AA index task | $7.76 at the $4/$20 standard price |
| API model ID | Not published |
Who can use it
Fairwind Program cyber defenders only, for now. Google is rolling Carbon out first to selected cyber defenders through its Fairwind Program, the vetted-access scheme it also used for Gemini 4 Argon, then to more Fairwind partners. It has not given a date for wider release (R&D World).
Why it is held back. Google says Carbon is its most capable model at finding and exploiting security holes, which is why defenders get it first.
Pricing
Google announced Carbon at $2 per million input tokens and $10 per million output tokens as an introductory price, with cached input at 95% off. After an introductory period Google has not dated, the price is $4 and $20. Artificial Analysis lists the $4/$20 standard price and uses it in its cost figures.
The introductory price matches Gemini 4 Argon’s. The standard price matches Claude Opus 5.5 ($4/$20).
Benchmark performance
Independent
| Measure | Gemini 4 Carbon | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| Intelligence Index v5.0 | 59 (high) | 58 (max) | 57 (max) |
| Cost per index task (standard price) | $7.76 | $10.04 | $6.25 |
| Output tokens per index task | ~123,000 | ~173,000 | n/a |
| Coding Agent Index | 59 | 62 | n/a |
| FrontierCode v1.1 | 51% | 56% | n/a |
Source: Artificial Analysis, with the GPT-6 Astra index score from its v5.0 index page.
Intelligence. 59 on the index at high reasoning, first on Artificial Analysis’s v5.0, which it launched the same day with two new benchmarks and most of the index on private datasets. Carbon is nine points above Gemini 4 Argon (50) at the same price. It is first on four of the nine evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA.
Factuality. Artificial Analysis scores it 54 on AA-Omniscience v1.1, against 48 for Opus 5.5 and 46 for Argon, with 58% accuracy and an 11% hallucination rate. Argon’s hallucination rate on the same version is 4%.
Coding. On Artificial Analysis’s runs, Carbon trails Opus 5.5 on coding: 59 to 62 on the Coding Agent Index and 51% to 56% on FrontierCode v1.1. It is second on Terminal-Bench 4.0 at 64%.
Cost per task. $7.76 per completed index task at the standard price, about 23% less than Opus 5.5 ($10.04) and about 1.2 times GPT-6 Astra ($6.25), on about 123,000 output tokens per task against Opus 5.5’s 173,000.
Knowledge work. On AA-Briefcase it scores 1,609 Elo, 126 above Argon, but below Kimi K3.1 Pro (1,678), Muse Wonder (1,659) and Grok 4.7 (1,638).
Google’s own figures
| Benchmark | Gemini 4 Carbon | Notes |
|---|---|---|
| SEC-bench Pro | 88.2% | Finding vulnerabilities; GPT-6 Astra 85.4%, Gemini 4 Argon 83.4% on Google’s table |
| ExploitGym | 54.8% | Exploiting vulnerabilities |
| CWE-bench | 71.0% | Patching vulnerabilities |
Source: Google, vendor-run, as reported by Interesting Engineering and Blockchain.News. Google ran these with guardrails disabled, so they describe the underlying model rather than the safeguarded product. It published no SWE-bench Pro or SWE-bench Verified figure.
How it compares
vs Claude Opus 5.5
Claude Opus 5.5 scores 58 on the v5.0 index, a point lower, at $4/$20 and $10.04 per task, and it is generally available. Opus 5.5 leads on coding (62 to 59 on the Coding Agent Index); Carbon leads on AA-Omniscience (54 to 48) and costs less per completed task.
vs GPT-6 Astra
GPT-6 Astra scores 57 on the v5.0 index at $6.25 per task. On Google’s own SEC-bench Pro table Carbon leads it, 88.2% to 85.4%.
vs Gemini 4 Argon
Gemini 4 Argon is generally available at the same $2/$10 introductory price and scores 50 on the v5.0 index, nine points under Carbon. Argon has the lower hallucination rate on AA-Omniscience v1.1 (4% to 11%).
Known limitations
Restricted. Carbon is only for selected cyber defenders in Google’s Fairwind Program, with no date for wider access.
No SWE-bench figure. Google published neither SWE-bench Pro nor SWE-bench Verified.
Introductory pricing. The $2/$10 price doubles to $4/$20 after an introductory period Google has not dated.
Documentation. No model card, knowledge cutoff, output limit or API model ID had been published at launch.
Frequently asked questions
When was Gemini 4 Carbon released?
Google announced it on 28 October 2026 and began rolling it out to selected cyber defenders in its Fairwind Program. It has not given a date for wider release.
Can I use Gemini 4 Carbon?
Not unless you are a cyber defender in Google’s Fairwind Program. It is not in the Gemini app or on the public Gemini API. Gemini 4 Argon is Google’s generally available frontier model.
How much does Gemini 4 Carbon cost?
$2 per million input tokens and $10 per million output tokens at the introductory price, with cached input at 95% off, then $4 and $20. Google has not dated the change.
Is Gemini 4 Carbon better than Claude Opus 5.5?
On Artificial Analysis’s Intelligence Index v5.0 it scores 59 to Opus 5.5’s 58 and costs less per completed task. Opus 5.5 is ahead on coding, at 62 to 59 on the Coding Agent Index, and it is generally available.
How is Gemini 4 Carbon different from Gemini 4 Argon?
Carbon scores nine points higher on the v5.0 index (59 to 50) at the same introductory price, and Google says it is much stronger at finding and exploiting security holes. Argon is generally available; Carbon is limited to Fairwind Program cyber defenders.