What is Gemini 4 Carbon?
Gemini 4 Carbon is Google DeepMind's new frontier model, announced on 28 October 2026, two weeks after Gemini 4 Argon reached subscribers. Google held back a general release because its cyber skills cut both ways: a model that finds and patches vulnerabilities can also exploit them. For now it goes only to a subset of cyber defenders in Google's invite-only Fairwind programme, and Google teams use it internally through Antigravity. Developers, businesses and consumers are promised it "as soon as possible", with no date.
There is no public API endpoint, model ID or model card yet. Artificial Analysis and Vals, which have tested it, list a 1M-token context window, and Vals lists 1,048,576 output tokens; it reads text, images, video, audio and files and writes text. Google has published no context or output limit of its own.
Google prices it like Argon: an introductory $2 per million input tokens and $10 per million output tokens, then $4 and $20, with cached input 95% off. When the introductory rate ends has not been announced.
How much does Gemini 4 Carbon cost?
| Plan | Price | What you get |
|---|---|---|
| Introductory | $2 in / $10 out per 1M tokens | Cached input $0.10; end date not announced |
| Later | $4 in / $20 out per 1M tokens | Cached input $0.20 |
Not yet on sale. Announced: $2 input / $10 output per 1M tokens introductory (cached input $0.10), then $4 / $20 (cached $0.20). End of the introductory rate not announced.
Is Gemini 4 Carbon worth it?
Carbon takes first place on the independent Artificial Analysis Intelligence Index v5.0 with 59, one point ahead of Claude Opus 5.5 (58), two ahead of GPT-6 Astra (57) and nine ahead of Argon (50). The margin is thin. Vals puts it second on its own index at 70.36%, behind Claude Sonnet 5.5 (70.85%) and just ahead of Opus 5.5 (70.18%), both within its ±0.97 margin. It does lead the Vals Finance Index and Legal Research Bench outright. This is a frontier model, not a runaway one.
Google's launch table is self-reported and has no error margins. It claims the top score on 15 of 17 rows, including 96.2% on GraphWalks long context against Astra's 71.8% and Opus's 66.8%. The cyber figures behind the restricted release, 88.2% on SEC-bench Pro for finding vulnerabilities and 54.8% on ExploitGym for exploiting them, were run without guardrails. NeuralTrust notes the ExploitGym run had a 10-hour cap against 6 hours for Opus 5.5; at 6 hours Carbon scores 53.5%. Even on Google's own table it trails on FrontierSWE 2 (61.2% against Astra's 65.5% and Opus's 62.3%) and on OSWorld computer use (81.5% against Opus's 87.2%).
Independent reliability data cuts against part of Google's story. Google says Carbon makes about half as many unsupported claims as Argon, yet per VentureBeat its hallucination rate on Artificial Analysis's Omniscience test rose from 4% to 11%. It answers more questions correctly (58% against 48%) but guesses more often when unsure. That is still far below the 50% recorded for Opus 5.5 and Astra.
It is slow and wordy. Artificial Analysis measures 118.58 seconds to first token against a median of 3.71, and 200 million output tokens to run its index against a median of 70 million. At the later $4/$20 price that comes to $7.76 per index task: close to Argon's $7.68, below Opus 5.5's $10.04 and above Astra's $6.25.
For anyone outside Fairwind, none of this is a buying decision yet. Per-token prices match Argon, so budgets set for Argon should carry over if Carbon reaches the Gemini API on the same terms, which Google has not confirmed. Teams on Opus 5.5 would gain a single index point from a model they cannot yet call.
Where it earns its place
A national CERT or critical-infrastructure security team admitted to Google's Fairwind programme
Find, validate and patch vulnerabilities in large codebases before attackers do
Google reports 88.2% on SEC-bench Pro and 71.0% on CWE-bench patching, both self-reported and run without guardrails, and says the model has run for days on internal engineering tasks. Access comes through a Google Cloud account team; per NeuralTrust, phishing-resistant MFA is required and use is limited to internal security, incident response or penetration testing teams.
A platform lead at a company already running Gemini 4 Argon in production
Budget for a move to Carbon once it opens up
Per-token prices match Argon, but Carbon spent 200 million output tokens on the Artificial Analysis index, so cost per task is the number to watch: $7.76 against Argon's $7.68 at the later rate. A job of 10M uncached input and 1M output tokens costs $30 at introductory prices and $60 later, per VentureBeat.
A head of AI at a bank or law firm comparing frontier models
Decide whether to wait for Carbon or commit to Opus 5.5 now
Carbon leads the Vals Finance Index at 66.35% and Legal Research Bench at 55.77%, but it is not on sale, and Opus 5.5 sits within the margin of error on the overall Vals Index. Waiting buys a small measured gain at an unknown date.
Before you commit
- Setup
- Not self-serve. Per NeuralTrust, Fairwind applications go through a Google Cloud account team, with background checks. Google has published no API model ID, and on 1 November 2026 the Gemini API models and pricing pages list only Argon from the Gemini 4 family.
- Integrations
- Inside Google it runs through Antigravity, and Varun Mohan, who leads Antigravity, named wider Antigravity access as the next step, without a date. No Gemini API, AI Studio, Vertex AI or Gemini app availability has been announced.
- Data and privacy
- Retention and training terms for Carbon are not published, and no Carbon system card exists yet. The Fairwind terms NeuralTrust describes (written for Argon) require usage tracking and forbid sharing or reselling access and creating malware; Carbon's may differ.
- Cost at scale
- At the later $4/$20 rate Artificial Analysis counts $7.76 per index task, with 123,000 output tokens per task; Vals records $10.51 per test on its index. Cached input at 95% off helps long shared prompts; whether Batch or Flex discounts will apply is not published.
- Limits
- Artificial Analysis and Vals list a 1M-token context and Vals 1,048,576 output tokens; Google has confirmed neither. Rate limits are not published. Artificial Analysis measures 118.58 seconds to first token and 84.6 tokens a second, and Vals took 47 minutes 45 seconds on its index.
What’s good
- First on Artificial Analysis Intelligence Index v5.0 at 59
- Leads the Vals Finance Index and Legal Research Bench
- Same per-token price as Argon; about 23% cheaper per index task than Opus 5.5
- 11% hallucination rate on AA-Omniscience, against 50% for Opus 5.5 and Astra
What’s not
- Only a subset of vetted cyber defenders can use it; no date for anyone else
- No API model ID, model card or Google-published limits
- 118-second median wait for the first token, and very verbose
- Trails Astra and Opus 5.5 on FrontierSWE 2, and Opus on OSWorld
Best for
Vetted security teams in Google's Fairwind programme today; Gemini API teams planning their next upgrade once Google opens access.
Gemini 4 Carbon alternatives
Gemini 4 Carbon: common questions
- Is Gemini 4 Carbon worth it?
- Carbon is the strongest model Google has built and, on Artificial Analysis's new index, the strongest anyone has, but almost nobody is allowed to use it yet. We rate it 4.3 out of 5.
- How much does Gemini 4 Carbon cost?
- Not yet on sale. Announced: $2 input / $10 output per 1M tokens introductory (cached input $0.10), then $4 / $20 (cached $0.20). End of the introductory rate not announced.
- What are the downsides of Gemini 4 Carbon?
- Only a subset of vetted cyber defenders can use it; no date for anyone else; no API model ID, model card or Google-published limits; 118-second median wait for the first token, and very verbose; trails Astra and Opus 5.5 on FrontierSWE 2, and Opus on OSWorld.
- Who is Gemini 4 Carbon best for?
- Vetted security teams in Google's Fairwind programme today; Gemini API teams planning their next upgrade once Google opens access.
- How much does Gemini 4 Carbon cost for a team or in production?
- At the later $4/$20 rate Artificial Analysis counts $7.76 per index task, with 123,000 output tokens per task; Vals records $10.51 per test on its index. Cached input at 95% off helps long shared prompts; whether Batch or Flex discounts will apply is not published.
- Is Gemini 4 Carbon safe for company data?
- Retention and training terms for Carbon are not published, and no Carbon system card exists yet. The Fairwind terms NeuralTrust describes (written for Argon) require usage tracking and forbid sharing or reselling access and creating malware; Carbon's may differ.
- What are the best Gemini 4 Carbon alternatives?
- The closest alternatives we have reviewed are Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon and GPT-5.6-Cyber.
Answers checked Nov 1, 2026.
Make Gemini 4 Carbon? Show this rating on your site
