Google DeepMind introduced Gemini 4 Carbon on 28 October, two weeks after Gemini 4 Argon became widely available, and calls it its strongest model so far. Carbon scores 59 on the independent Artificial Analysis Intelligence Index, one point above Claude Opus 5.5. For now only a selected group of cyber defenders gets access, and Google gives no date for developers, businesses or consumers.
There is no Carbon in the Gemini app, the Gemini API or Gemini Enterprise yet. Google first wants feedback from those defenders and time to adjust its safety measures. The model is also in the US government’s voluntary testing of new models before release.
Carbon’s biggest gain is in turning software flaws into working attacks
Google tests cyber skills in three steps: finding a vulnerability, checking whether an attacker can exploit it, and writing a patch. Carbon improved on all three, but the largest jump is in the middle step. That explains the order of access. A model that builds working attacks more often is most useful to whoever gets it first, and Google wants that to be the people who patch.
Argon is the model you can use today, at the same introductory price, so it is the fair comparison.
| What is measured | Gemini 4 Carbon | Gemini 4 Argon |
|---|---|---|
| Overall score, Artificial Analysis Intelligence Index v5.0 | 59 | 50 |
| Finding vulnerabilities (SEC-bench Pro) | 88.2% | 83.4% |
| Turning vulnerabilities into working attacks (ExploitGym) | 54.8% | 42.0% |
| Writing patches (CWE-bench) | 71.0% | 68.0% |
| Coding agents, Artificial Analysis Coding Agent Index | 59 | 53 |
The two index scores come from Artificial Analysis, an independent benchmarking company. The three security scores are Google’s own, measured without the safety filters customers get, and nobody outside Google has repeated them yet.
Google also says Carbon did more than ten times better than an unnamed rival model at writing good patches for the Chrome team, but shows no data for it.
Claude Opus 5.5 still leads on coding agents
Carbon’s lead on the Artificial Analysis index is narrow: 59 against 58 for Claude Opus 5.5 and 57 for GPT-6 Astra. It is also measured on version 5.0, which Artificial Analysis rebuilt with harder tests, so older scores do not compare. On the same company’s coding agent index, Claude Code with Opus 5.5 scores 62 and Carbon 59. On knowledge work such as reports and analyses, Kimi K3.1 Pro, Meta’s Muse Wonder and Grok 4.7 score higher.
Is Gemini 4 Carbon more reliable than Argon?
On Google’s own measures it is, but an independent test points the other way. Google says Carbon makes about half as many unsupported claims as Argon. It also stopped every attack in Gray Swan’s prompt injection test, where hidden instructions in a web page or document try to take over the model.
Artificial Analysis measured a related habit with factual questions. When Carbon did not get a question right, it gave a wrong answer instead of a partial one or “I don’t know” in 11% of cases. For Argon that was 4%. Claude Opus 5.5 and GPT-6 Astra both score 50% on this test, so Carbon still guesses far less than they do, just more than its predecessor.
Carbon starts at Argon’s price, and Google has not said for how long
Once Carbon opens up, it will cost $2 per million input tokens and $10 per million output tokens, with 95% off for cached input. That is Argon’s introductory price. Afterwards it doubles to $4 and $20. Argon’s introductory period ends on 30 November, but Google gives no end date for Carbon.
The same price per token does not mean the same bill. Carbon works longer on hard tasks, and The Decoder reports that some finance and tax agent tasks take more than twice as long as with Argon. Compare the cost per finished task in your own test, not the price list.
Can Dutch organisations get access to Gemini 4 Carbon?
Only through Google’s Fairwind Program, and only if you defend important systems. Fairwind gives priority to governments and national cyber authorities, operators of critical infrastructure such as healthcare, telecom, energy and financial networks, and core technology platforms. Academic labs that work on defensive benchmarks can apply as well. Google vets every applicant, and you apply through your Google Cloud account team, as the Fairwind Program page explains.
The Fairwind page names no regional limits, and Google’s announcement does not mention the EU or the Netherlands. Only internal security, incident response and penetration testing teams may use the model, and they must sign in with phishing-resistant multi-factor authentication. The page still describes Argon and does not mention Carbon yet.
Use the wait for Carbon to shorten your patch cycle
For most organisations DataNorth works with, Carbon changes nothing this month. You cannot buy it, and Argon remains the Google model to test. What does change is how fast software flaws are found. In the PHP2Shell case last week, Gemini 4 Argon and GPT-5.6 Sol helped security firm Wiz turn PHP bugs into working attacks. Dutch hospitals and the Tweede Kamer took sites offline until they could patch. Carbon does that step better, and Google plans to release it to developers, businesses and consumers.
That makes this a good moment to tighten your own vulnerability management:
- Keep a list of the software versions on every system reachable from the internet, including the ones your suppliers manage.
- Agree on patch times in days, not in quarterly maintenance windows, for those systems.
- If you build software, use a current model as a security reviewer on your own code now. Carbon can replace it later without changing the process.
- If you run critical infrastructure, ask your Google Cloud account team about Fairwind.
Teams that use Gemini for writing, summaries or analysis lose little by waiting. Carbon does not lead those tests, and Argon is available at the same price today.
For more information, visit the official announcement of Gemini 4 Carbon on the Google blog.
