Google has introduced Gemini 4 Carbon, a new frontier model it describes as its strongest to date, with a pronounced focus on coding and cybersecurity. The announcement came in a Google blog post by Koray Kavukcuoglu dated October 28, 2026, and was covered by Engadget on October 29. Carbon arrives about two weeks after Gemini 4 Argon.
The twist is who can use it. Google is not releasing Carbon to developers, businesses or consumers yet. For now it is limited to a subset of cyber defenders, and the company says access will widen to trusted partners through a programme called Fairwind.
What Google claims
On the Artificial Analysis Intelligence Index v5.0, Carbon ranks first with a score of 59, according to Google. WinBuzzer, which reported the index independently, lists Anthropic's Opus 5.5 at 58, OpenAI's GPT-6 Astra at 57 and Gemini 4 Argon at 50. A one-point lead is narrow, and the gap to the next model is within what many readers would call a statistical tie.
Security results are where Google makes its biggest claims. Carbon scores 88.2 percent on SEC-bench Pro, a test of finding vulnerabilities, and 54.8 percent on ExploitGym, which measures exploitation, up from 42.0 percent for Argon according to WinBuzzer. On CWE-bench, a test by Collinear of patching root causes, it reaches 71.0 percent, against 68 percent for Argon and 67 percent for Opus 5.5. Google also reports an indirect prompt-injection attack success rate of 0.0 percent in a Gray Swan test, compared with 0.7 percent for Argon.
For coding, Google reports 65.2 percent on Terminal-Bench 4.0, up from 57.1 percent, and says attempts to cheat the test dropped to 0.4 percent, with no successful cheating. It places fourth on WebDev Arena with a rating of 1768, an improvement of 89 points. Google says the model more than doubles secure-code generation compared with Argon, but that figure comes from internal testing.
Google also quotes partners. The Chrome team says Carbon produces high-quality patches more than ten times as often as another lab's model, and security firm Wiz says it leads on web penetration-testing benchmarks. Google says the external cyber benchmarks were run without guardrails.
Where it falls short
Engadget describes Google as claiming a near-sweep, but WinBuzzer's reporting shows gaps. On FrontierSWE 2, Carbon scored 61.2 percent, behind GPT-6 Astra at 65.5 percent and Opus 5.5 at 62.3 percent. On an offline subset of OSWorld, Opus 5.5 reached 87.2 percent against 81.5 percent for Carbon. The same report cites an independent hallucination rate of 11 percent for Carbon, compared with 4 percent for Argon. Treat any claim that Carbon wins across the board as the company's framing.
Pricing and access
Google lists introductory pricing of $2 per million input tokens and $10 per million output tokens, with a 95 percent discount on cached input. After the introductory period the price rises to $4 and $20. Google has not said when the introductory period ends, and Engadget said pricing was not mentioned in its piece, so the figures come from Google's blog and WinBuzzer.
Fairwind, according to WinBuzzer, covers more than 650 partners, requires background checks and phishing-resistant multi-factor authentication, and restricts use to internal security, incident response and penetration-testing teams. Partners may not share or resell access. Google says a broader release will come "as soon as possible" but gives no date. It has also joined the US government's voluntary pre-release testing programme for frontier models.
Why the limited release matters
Releasing a model this good at finding and exploiting software flaws to everyone at once would also hand those abilities to attackers. By starting with defenders, Google is trying to give patching teams a head start. Critics will ask how well a vetting scheme holds up at the scale of hundreds of partners, and whether a competitor will release similar capabilities without restrictions.
For Indian developers and security teams, nothing has changed yet, since Google has announced no regional availability, and the company has not said whether Indian firms can join Fairwind.
What to watch next
The key signals will be the date Google opens Carbon to general developers, when the introductory price ends, and independent evaluations of its higher hallucination rate. Rival labs are likely to respond to the cyber benchmarks with their own results.
Published October 30, 2026. Gadgets365 will update this article as more information becomes available.
