Google's Gemini 4 Carbon leads benchmarks with a cyber-defence focus
Google's new frontier model tops key AI and security leaderboards, but only vetted defence firms can use it for now. Here is what it means for India.
Key points
- Google has released Gemini 4 Carbon, a frontier model it says leads the Artificial Analysis Intelligence Index (v5.0).
- Access is limited to select cybersecurity defence firms for now, priced at $2 per million input tokens and $10 per million output tokens.
- Indian security teams and developers may benefit once access widens, but dollar pricing leaves rupee costs exposed to currency moves.
The News
Google has put out Gemini 4 Carbon, its newest frontier model, and is pitching it first and foremost as a tool for cyber defence. The company announced the release on Wednesday and said the model now ranks first on the Artificial Analysis Intelligence Index (v5.0), placing ahead of Claude Opus 5.5 and GPT-6 Astra on that leaderboard.
Access is restricted for now. Only a select group of cybersecurity defence companies can use Carbon today, with Google planning a wider rollout to other partners in its Fairwind Program later. The model succeeds Argon, which Google shipped earlier this month.
On security-specific tests, Carbon holds the top position on SEC-bench Pro, which checks how well a model uncovers real-world bugs, and on ExploitGym, which measures whether a model can turn a known vulnerability into a working exploit. Google also says the model cuts "unsupported claims" by two times compared with Argon, and that internal teams have watched it keep working for days on long, multi-step engineering jobs.
Pricing is unchanged from Argon: $2 per million input tokens and $10 per million output tokens, described by Google as an introductory rate.
Why It Matters
Two releases in one month signals how compressed the frontier-model cycle has become. When OpenAI launched GPT-4 in March 2023, rivals needed many months to answer it. Google now appears willing to replace its own flagship within weeks, and to hold the price flat while doing so, which puts pressure on competitors' margins as much as on their benchmark scores.
The gated launch is the more telling choice. A model that tops an exploit-generation benchmark is useful to defenders and attackers alike, so restricting early access to vetted security firms is a way of letting the defensive side get a head start. Expect other labs to adopt similar staged releases for models with strong offensive-security capability.
Indian Angle
India's security market is the obvious place to watch. Under CERT-In's April 2022 directions, organisations must report specified cyber incidents within six hours, and banks and NBFCs face their own RBI cyber-resilience expectations. Tools that find vulnerabilities and write safer code faster could help security teams at Indian lenders, payment firms and IT services majors meet those obligations, though only once Carbon moves beyond its initial partner group.
The flat pricing matters for Indian developers who pay in dollars. Because the rate is quoted in US currency, rupee depreciation raises the effective cost even when the list price stays at $2 and $10 per million tokens. Startups building agentic coding products on top of Gemini can budget against an unchanged dollar rate, but should model currency risk into their unit economics.
For domestic model builders such as Sarvam and Krutrim, Carbon raises the bar again. Competing head-on with frontier labs on general benchmarks is difficult, which strengthens the case for Indian firms to specialise in Indic languages, sovereign deployments and sector-specific models where data residency under the Digital Personal Data Protection Act, 2023 is a selling point.
FAQ
Can anyone use Gemini 4 Carbon today?
No. Google has limited access to a select group of cybersecurity defence companies. It plans to expand availability to other partners in its Fairwind Program over time, but has not given a date for wider public access. When it does open up, pricing will match Argon's introductory rate.
How much does Gemini 4 Carbon cost?
Google is charging $2 per million input tokens and $10 per million output tokens, the same introductory price it set for Argon. Indian users billed in dollars should note that their rupee cost will move with the exchange rate even though the list price has not changed.
How does Carbon compare with Claude Opus 5.5 and GPT-6 Astra?
According to Google, Carbon ranks first on the Artificial Analysis Intelligence Index (v5.0), with Claude Opus 5.5 and GPT-6 Astra behind it. It also leads SEC-bench Pro and ExploitGym. These are benchmark results, so real-world performance on a given workload may differ.
What does this mean for Indian security teams?
If access widens, Carbon could help Indian security teams find bugs and write more secure code faster, supporting compliance with CERT-In's six-hour incident reporting rule and RBI cyber-resilience norms. For now, only firms in Google's initial partner group can use it.
This story was reported by The Verge. Read the full original coverage at The Verge.