What happened
- On October 28, 2026, Google announced Gemini 4 Carbon, its new frontier model, with gains in cybersecurity and in long programming tasks.
- The company reported scores of 88.2% on SEC-bench Pro, which measures vulnerability detection; 54.8% on ExploitGym, which measures exploitation; and 71.0% on CWE-bench, which measures the ability to fix flaws.
- Carbon arrives two weeks after Gemini 4 Argon and, according to Google, tops version 5.0 of the Artificial Analysis Intelligence Index with 59 points.
- Starting this week, the model will be available only to a group of cyber defenders, and later to trusted partners through the Fairwind program.
Why it matters
- A model that finds, exploits and fixes flaws can serve defense as well as attack. That is why Google chose to give it first to defense teams and to take part in the US government’s voluntary pre-release testing process.
- Ars Technica reports that, according to Google, Carbon produces vulnerability-free code twice as often as Argon, and that the Chrome team rates it ten times better at producing quality patches than models from other companies.
- The company also says that, in real programming environments with its own teams, Carbon showed lower rates of misalignment and fabrication. These are internal measurements, with no public detail of the method.
- For companies in Latin America there is no immediate access: the model will not be in developer or consumer products until Google finishes tuning its safeguards.
The number
88.2%. That is the score Google reports for Gemini 4 Carbon on SEC-bench Pro, the vulnerability-finding test.
Context
Google describes Carbon as a model built for sustained agent workflows. It says it has seen it work internally for days on multi-step tasks, such as tracking down the cause of subtle bugs in large codebases and fixing them.
Ars Technica tempers the announcement. It notes that Carbon beats Argon by slim margins in each test and that the gaps with OpenAI and Anthropic models are a few points, as happens across all frontier models. It also points out that Google presents it as a new model, although it could be a post-training upgrade of Argon. According to the outlet, the model resisted every prompt-injection attempt in the Grey Swan evaluation.
When it opens, Carbon will carry the same launch price as Argon: US$ 2 per million input tokens and US$ 10 per million output tokens, with a 95% discount on cached input.
What’s next
- Google gave no date for general access; it says it will keep gathering feedback and iterating on guardrails before opening it.
- The Fairwind program will be the entry point for trusted partners in the coming weeks.
Bottom line
The race for flaw-finding models goes back a while: Claude verified the security of 123 open libraries and an open Chinese model nearly matched recent closed models on exploits. In practice, PHP fixed two remote execution flaws stalled for two years, and METR and Redwood found ways around OpenAI’s agent monitor.
Sources
- Google — Gemini 4 Carbon, our next frontier model (October 28, 2026)
- Ars Technica — Google announces Gemini 4 Carbon, its latest frontier AI model that you can’t use (October 28, 2026)
- Cover image: BalticServers.com, CC BY-SA 3.0, via Wikimedia Commons
Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.
