Tech • AI • Robotics • Game
Google Just Announced Gemini 4 Carbon (It Cooks)
9.4/10TL;DR
Google has unveiled Gemini 4 Carbon, a new flagship AI model positioned above Gemini 4 Argon, with claimed strengths in cybersecurity, long-running autonomous work and benchmark performance, though access is initially restricted.
KEY POINTS
A rapid follow-up to Argon
Gemini 4 Carbon arrives only two weeks after Gemini 4 Argon, signaling a much faster release cadence from Google in the top-tier AI race. The launch suggests the company is now iterating at a pace closer to leading rivals rather than spacing major model updates months apart.
Built for long-horizon tasks
Google describes Carbon as a model designed for long-horizon work, meaning complex assignments that require many steps and extended execution time. Internal testing reportedly showed the model working for days on difficult engineering tasks while continuing past dead ends instead of stopping early, a capability that could expand what organizations can delegate to AI agents.
Cybersecurity is the headline use case
The strongest emphasis is on cybersecurity. Carbon is designed to identify software vulnerabilities, verify that they are genuine, exploit them in controlled testing and then help patch them, making it relevant both for defensive security teams and for software maintenance workflows.
Benchmark lead over major rivals
On the results shared by Google, Carbon outperformed GBT6 Astra, Claude Opus 5.5 and Claude Fable 5.1 on 16 of 18 benchmarks. The only exceptions were one coding benchmark led by GBT6 Astra and one computer-use test led by Claude Opus 5.5. Carbon also exceeded Argon on every benchmark disclosed.
Strong security-specific scores
Carbon led all three of Google’s security evaluations, covering vulnerability discovery, exploitation and patching. On Grey Swan’s prompt-injection benchmark, which measures whether hidden malicious instructions can manipulate the model, Carbon posted a 0% attack success rate, indicating no successful prompt-injection attempts in that test set.
Not best in every category
Despite its overall lead, Carbon is not dominant across all domains. In web development, it placed fourth on WebDev Arena, suggesting that while it is competitive, it is not the strongest model currently available for every practical coding workload.
Lower hallucination and no cheating in tests
Google says Carbon makes about half as many unsupported claims as Argon, a sign of improved reliability. In coding-related evaluations aimed at detecting deceptive behavior, the model reportedly scored 0% on cheating, meaning it did not fabricate successful results to appear as though tasks were completed.
Staged rollout before public release
Carbon is not yet generally available. Beginning this week, access is being granted first to a small group of cyber defenders, then to a limited set of trusted partners. The strategy mirrors a broader industry practice in which advanced models are first placed with defenders so vulnerabilities can be patched before wider distribution.
Public timeline and pricing
No public release date has been announced, but Argon followed a similar path, moving from security-first access on September 30 to broader public availability on October 13. For developers, Carbon will debut at the same introductory pricing as Argon: $2 per million input tokens and $10 per million output tokens. After the introductory period, pricing is expected to rise, though the stated figure of $420 appears unclear as presented.
Questions remain over naming and lineup
The emergence of Argon and Carbon leaves open questions about how Google intends to structure the Gemini 4 family. Carbon is clearly presented as the stronger successor, but the broader hierarchy of future models has not yet been spelled out.
CONCLUSION
Gemini 4 Carbon strengthens Google’s position in the frontier AI race with aggressive benchmark claims and a sharp focus on cybersecurity. Its real impact will depend on whether those results hold up once the model moves from restricted defensive use to wider public access.
Ask a question