Google DeepMind released Gemini 4 Carbon on October 28, 2026, two weeks after Gemini 4 Argon. The model ranks first on the Artificial Analysis Intelligence Index v5.0, but access is restricted to selected cybersecurity experts, which limits who can use it in practice.
- Gemini 4 Carbon arrived on October 28, 2026, two weeks after Argon, and it ranks first on Artificial Analysis Intelligence Index v5.0 with 59 points.
- Access starts with selected cybersecurity experts, then widens through the Fairwind Program. No date is given for general release.
- Carbon leads all tested models on SEC-bench Pro, takes first place on CWE-Bench by a wide margin and also posts a result on ExploitGym.
- Introductory pricing is $2 per million input and $10 per million output tokens, with no stated end date.
- Accuracy rose. So did the hallucination rate.
Google releases Gemini 4 Carbon to cybersecurity experts first
Google DeepMind launched Gemini 4 Carbon on October 28, 2026, according to The Decoder. The release came two weeks after Gemini 4 Argon, so the company has put two top-tier models out in a short span.
Google calls Carbon its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities, according to apidog.com. The announcement was made by Koray Kavukcuoglu.
Initially, Carbon is available only to selected cybersecurity experts, starting the week of the announcement, Yahoo Tech reports. The restricted rollout is tied to the model’s advanced security capabilities.
Carbon is also going through the U.S. government’s voluntary pre-testing process. The sources do not describe what that testing involves or when it ends.
For teams that build workflows on Google models, including content and SEO operations, the practical result is simple: Carbon can be read about today but not used, and nothing in the sources says when that changes.
Carbon takes first place on the Artificial Analysis index
The Artificial Analysis Intelligence Index is an independent benchmark that ranks AI models; the current version is 5.0. Carbon scores 59 points and takes first place, while Argon scores 50.
The comparison needs a caveat. Artificial Analysis listed Argon (High) at 53 on October 15, before the v5.0 update. Part of the gap between Carbon and Argon therefore reflects the changed index and not only the new model.
The Decoder says Carbon at its high reasoning tier leads mainly because of top scores in complex reasoning, knowledge retrieval and scientific tasks. It takes first place in four of nine individual evaluations, which means it does not win across the board. Among the Gemini 4 Carbon benchmark results, AA-Briefcase is one more Artificial Analysis evaluation where Carbon improves on Argon.
Coding and arena results
On the Artificial Analysis Coding Agent Index, Carbon scores 59 against Argon’s 50. It still trails Claude Opus 5.5, which scores 62.
On Arena.ai, Gemini 4 Carbon (High) takes the Text Arena lead with 1,528 points, ahead of Argon at 1,520. The margin is eight points.
On WebDev Arena the model gained 89 Elo points to reach 1,769, good for fourth place. Elo is a rating scheme in which points rise or fall with head-to-head results. The Decoder gives 1,768 in one place and 1,769 in another for this figure, so the exact number is slightly uncertain.
Accuracy and agent rankings
On Vals AI, Carbon ranks second in accuracy. Claude Sonnet 5.5 leads narrowly, and Opus 5.5 follows just behind Carbon. The three scores are less than one percentage point apart. In the Vals overall index, Carbon also places second among the models tracked there, so the independent results point the same way. For content teams comparing these models, Anthropic’s watermarking of Claude-generated text is another factor worth weighing alongside accuracy scores.
In Agent Arena, Carbon places second across more than 21,900 real sessions, ahead of GPT-6 Astra and Fable-5.1. Because the gaps in several rankings are narrow, positions could shift with further evaluations.
Taken together, the results show a model that leads on the aggregate index and in text, but that sits behind rivals in coding agents and in web development.
Security benchmarks show the model’s main strength
SEC-bench Pro measures autonomous vulnerability discovery in real-world software. Carbon leads all tested models there. Google ran it without safety guardrails for this test.
ExploitGym tests whether a system can turn discovered vulnerabilities into working exploits. Carbon’s score there is well below its vulnerability-discovery result.
On CWE-Bench, which covers patching, Carbon takes first place with a wide lead. That result comes from the Yahoo Tech report; The Decoder reports no number for it.
Read together, the three scores cover finding flaws, exploiting them and fixing them. That range explains why Google chose security specialists as the first group to get access.
Resistance to manipulation
Gray Swan’s indirect prompt injection benchmark tests whether a model resists instructions hidden in content it reads. Carbon cut the attack success rate, already low for Argon, to zero.
On Terminal-Bench 4.0, successful cheating fell to zero, and cheating attempts became rare.
Limits of the evidence
Google reports that Carbon produces secure code more than twice as often as Argon on in-house secure-code tests. These internal results cannot be independently verified.
What partners and Google’s own teams report
The Chrome team reported that Carbon performed many times better than another lab’s frontier model at generating high-quality security patches. The comparison model is not named.
Cloud security platform Wiz observed Carbon leading on benchmarks for web penetration testing.
Google says internal teams ran Carbon for multiple days on long-horizon engineering tasks, such as root-cause analysis and subtle bug fixes in large codebases.
Early testers cited by TestingCatalog, as relayed by apidog.com, say Carbon beats Argon at 3D work and is weaker at frontend and UI. They also describe it as slower but less lazy.
These accounts come from Google, a partner and early testers, not from independent controlled tests, so they read as indications and not as measured results.
Accuracy improves while hallucinations rise
In AA-Omniscience, the Artificial Analysis knowledge evaluation, Carbon’s answer accuracy rises well above Argon’s. Its hallucination rate, meaning how often it states something false instead of declining to answer, also rises well above Argon’s much lower rate.
The Decoder notes that Carbon’s hallucination rate is still the second-lowest in the ranking. Opus 5.5 (max) and GPT-6 Astra (max) both trail on answer accuracy, so Carbon is a clear step ahead of both.
The trade-off is concrete: the model answers more questions correctly, but when it is wrong it is more likely to give a confident false answer than Argon was.
Google links the Gemini 4 safety report, which covers Argon’s alignment results, and says Carbon uses the same multi-layer safety stack. The report therefore does not present separate alignment results for Carbon.
Pricing and token use compared with rivals
Carbon’s introductory price is $2 per million input tokens and $10 per million for output, matching Argon’s introductory price. Cached input is priced far below regular input. Tokens are the chunks of text a model reads and writes, and API usage is billed by them.
Artificial Analysis bases its calculations on a list price of $4 per million input tokens and $20 for output. It ignores the current introductory discount that runs through the end of 2026. The end date of Carbon’s own introductory period has not been stated.
| Model | Cost per task in the index |
|---|---|
| Gemini 4 Carbon | $7.76 |
| Claude Opus 5.5 (max) | $10.04 |
| GPT-6 Astra (max) | $6.25 |
Carbon is cheaper per task than Claude Opus 5.5 but costs more than GPT-6 Astra on this measure. It averages 123,000 output tokens per task, far fewer than the 173,000 that Anthropic’s model uses.
Against Argon, though, it generates far more reasoning tokens and considerably more output tokens, so token use per task is higher than with the earlier model.
How Carbon relates to Gemini 4 Argon
The table sums up how the two models compare on the figures reported so far.
| Measure | Gemini 4 Carbon | Gemini 4 Argon |
|---|---|---|
| Artificial Analysis Intelligence Index v5.0 | 59 points | 50 points |
| Coding Agent Index | 59 | 50 |
| Text Arena | 1,528 | 1,520 |
| WebDev Arena | 1,769 (fourth place) | About 89 points lower |
| Hallucination rate | Higher | Lower |
| Access | Selected cybersecurity experts | Stable on the paid Gemini API |
Argon is Google’s flagship reasoning model for developers. Its model ID is gemini-4-argon, and it became stable on the paid Gemini API on October 13. It has roughly a 1M-token unified context window.
Argon’s introductory price of $2 per million input and $10 per million output tokens lasts through November 30. List prices of $4 and $20 apply from December 1.
Carbon, by contrast, is not yet an API product for most users, and none of those details have been published for it.
The name Carbon was in circulation before the October 28 announcement. The model was first reported as an internal checkpoint at Google, which explains why it showed up in reports ahead of the launch. Carbon belongs to the same family of internal codenames as Argon, and both models kept those names when Google made them public. For readers who saw the name earlier, that history connects those reports to the model released now.
Who gets access and what is still unknown
The Fairwind Program is Google DeepMind’s early-access initiative for trusted partners. After the first rollout to cyber defenders, it expands access to additional trusted partners.
Google plans to open Carbon to developers, businesses and consumers through its product interfaces after collecting more feedback and adjusting safety guardrails. No date has been given.
As of October 29, 2026, Carbon’s model ID, context window and API date had not been published. Teams planning around Carbon have no confirmed timeline for access, while Argon remains the Gemini 4 model available on the paid API.
Sources
Photo: The Decoder
Frequently asked questions
What is Gemini 4 Carbon?
What is Gemini 4 Carbon?
It is a model from Google DeepMind that Google calls its strongest, with state-of-the-art cybersecurity and long-horizon capabilities. It was released on October 28, 2026.
How does Gemini 4 Carbon score on the Artificial Analysis Intelligence Index?
How does Gemini 4 Carbon score on the Artificial Analysis Intelligence Index?
It scores 59 on version 5.0, ranking first. Argon scores 50 on the same version.
Who can use Gemini 4 Carbon now?
Who can use Gemini 4 Carbon now?
Only selected cybersecurity experts, starting the week of the announcement. More trusted partners will follow via the Fairwind Program.
How much does Gemini 4 Carbon cost?
How much does Gemini 4 Carbon cost?
The introductory price is $2 per million input tokens and $10 per million output tokens, with cached input 95 percent cheaper. The end date of the introductory period has not been stated.
Is Gemini 4 Carbon better than Gemini 4 Argon?
Is Gemini 4 Carbon better than Gemini 4 Argon?
Carbon scores higher on the index (59 versus 50) and in accuracy (58 versus 48 percent). Its hallucination rate is higher, at 11 percent against 4 percent, and early testers call it slower.
When will Gemini 4 Carbon be available to everyone?
When will Gemini 4 Carbon be available to everyone?
No date is given. Google plans to open it to developers, businesses and consumers after gathering feedback and adjusting safety guardrails.