The model improves vulnerability detection and software patching, with a phased rollout and introductory pricing matching Gemini 4 Argon
Google is introducing Gemini 4 Carbon through a phased rollout beginning with selected cybersecurity defenders.
Google is rolling out Gemini 4 Carbon first to selected cybersecurity defenders, giving them access to its latest AI model for finding software vulnerabilities, checking whether they can be exploited and producing fixes.
Access will expand through Google DeepMind’s Fairwind Program before a broader release to developers, businesses and consumers. Google describes Carbon as its strongest model, with improvements in reasoning and engineering tasks that require sustained work across multiple steps.
Carbon debuted at number one on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59, according to Google’s launch announcement.
Google DeepMind’s Clément Farabet discussed the model on LinkedIn: “Beyond writing new code, Carbon shows a major leap in defensive cybersecurity—reliably tracing subtle vulnerabilities across large, interconnected codebases and producing verified root-cause patches.”
Finding vulnerabilities and fixing their causes
Google reports that Carbon scored 88.2 percent on SEC-bench Pro, which tests its ability to uncover vulnerabilities in real software, and 54.8 percent on ExploitGym, which assesses whether it can demonstrate that vulnerabilities are exploitable.
That second step helps security teams distinguish between a potential weakness and one that could be used to compromise a system, informing which fixes they prioritize.
Carbon also reached 71 percent on Collinear’s CWE-bench, which tests its ability to repair vulnerabilities. Google evaluated these cybersecurity benchmarks without guardrails, the restrictions used to control a model’s behavior.
Within Google, Carbon more than doubled Argon’s success rate on internal tests of secure code generation. The company says its engineering teams have also seen the model work for days on difficult tasks, investigating dead ends and tracing bugs across large software projects.
Those capabilities extend beyond security work. Google reports improvements in mathematical reasoning, scientific tasks and web development, with Carbon reaching fourth place on WebDev Arena at launch.
Introductory pricing matches Argon
Carbon will initially cost $2 per million input tokens and $10 per million output tokens, matching Gemini 4 Argon’s introductory rates. Tokens are the units of information a model processes, with input covering material it receives and output covering what it generates.
Cached input tokens, which allow previously processed material to be reused, will receive a 95 percent discount on the input rate.
Google’s pricing footnote states that rates will rise to $4 per million input tokens and $20 per million output tokens after the introductory period.
Farabet highlighted the relationship between performance and operating costs in his LinkedIn post, describing Carbon as delivering stronger reasoning without an increase over Argon’s token pricing.
Controlled access and stronger safeguards
Google is gathering feedback and refining safeguards through the initial rollout. It is also participating in the US government’s voluntary process for testing models before release.
The Fairwind Program prioritizes organizations defending critical systems, including governments, healthcare providers and telecommunications services. Google DeepMind also welcomes applications from academic laboratories focused on defensive benchmarking, making the program relevant to university cybersecurity research.
Alongside Carbon’s stronger capabilities, Google reports improvements in how it follows instructions and handles uncertainty. Internal testing found roughly half as many unsupported claims as Argon, while high-severity misalignment alerts fell by almost 80 percent in live coding environments used by Google teams.
Carbon also recorded a zero percent attack success rate on the evaluated Gray Swan indirect prompt injection benchmark. These attacks try to redirect an AI system through malicious instructions hidden in material it reads, such as documents or web pages.
Farabet wrote: “Given its strength in vulnerability discovery and exploit analysis, Carbon is rolling out first to cybersecurity defenders and trusted partners through the Fairwind Program before opening up more broadly to developers.”
ETIH Innovation Awards 2027 open for entries
The ETIH Innovation Awards 2027 are now open for entries, with 26 categories recognising companies, institutions, projects and individuals making a measurable impact across education technology. The global awards span AI in education, digital learning, higher education, skills and workforce, inclusion, education operations, evidence and impact, with entries managed through ETIH’s new online awards platform.
Entries are open worldwide, with the early-bird deadline on Friday 18 December 2026 and the final entry deadline on Friday 26 February 2027 at 23:59 UK time.
Explore the 2027 categories and judging criteria, read the How to Enter guide or start an entry through the ETIH Innovation Awards 2027 platform.
