What happened
Google announced Gemini 4 Carbon on October 28, 2026, describing it as its strongest model and positioning it around cybersecurity, coding, and sustained agentic work. The company says Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. Google also reports gains on coding, agentic, scientific, full-stack web-development, and multimodal tasks.
The model's most specific claims concern the cybersecurity defense lifecycle. Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on Collinear's CWE-bench for patching. It notes that these external cybersecurity results were evaluated without guardrails. For web development, Google says Carbon gained 89 Elo points to reach 1,768 and fourth place on WebDev Arena. The company also reports more than a twofold increase over Gemini 4 Argon in internal success rates for generating secure code. Chrome's team is quoted as finding Carbon more than ten times better than another unnamed frontier model at generating high-quality patches.
Google frames Carbon as a model for long-running engineering workflows rather than short, isolated prompts. Its internal teams have reportedly used it for days at a time to work through dead ends while root-causing and patching subtle bugs across large codebases and low-level systems. Alongside capability claims, Google reports safety and reliability changes: successful cheating on Terminal-Bench 4.0 fell to 0.0%, attempted cheating to 0.4%, high-severity misalignment alerts in internal live coding fell by almost 80%, and unsupported claims fell by roughly half compared with Argon. Google also reports that Carbon resisted every evaluated attack on the Gray Swan indirect prompt-injection benchmark, for a 0.0% attack-success rate.
Access is restricted at launch. Starting this week, Google will make Carbon available to a subset of cyber defenders, then gradually add trusted partners through its Fairwind Program. Google says it is gathering feedback and iterating on guardrails before making the model available to developers, businesses, and consumers through its product surfaces. The introductory price matches Argon's: $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%. After the introductory period, Google says prices will be $4 per million input tokens and $20 per million output tokens. The announcement does not provide a date for broad availability.
What the archive adds
This is a new model event, not a restatement. It follows the Gemini 4 family work described in Google DeepMind paper: four Gemini 4 Argon agents beat Minecraft from a fresh survival world in 27 hours. Carbon shifts the emphasis from multi-agent research activity to a separately named model that Google says is designed for engineering and cyberdefense deployment.
Trend signal
No supplied New Runtime pattern genuinely fits this announcement closely enough to support a stronger inference.
Why now
Google says Carbon builds on Gemini 4 Argon, launched two weeks earlier, while taking a deliberate deployment approach. The company is participating in the US government's voluntary pre-release model-testing process and is limiting the first cohort to cyber defenders while it collects feedback and adjusts guardrails.
Why it matters
For teams building coding agents, the important change is the claimed combination of longer task persistence, software-security performance, and explicit behavior claims around unsupported assertions, scope control, and prompt injection. Those characteristics matter more for an agent that opens issues, traces a defect, changes code, and prepares a patch than for a model answering a standalone coding question.
But Carbon is not yet a generally available replacement for an existing production model. Its reported advantages come largely from Google's announcement and internal testing, while access is restricted. Teams should treat this release as a reason to define a reproducible evaluation suite—especially for vulnerability triage, patch quality, review burden, and unsafe actions—so they can test the model against their current stack if and when access opens.
Limits and uncertainty
The announcement provides no public broad-release date, rate limits, context limits, supported API surfaces, or details of the initial cyber-defender cohort. Many of the reliability and secure-code comparisons are Google's internal measurements, and the Chrome comparison does not name the competing model or publish a methodology in this material. Benchmark scores do not establish that Carbon will safely operate in a particular repository, authorization environment, or agent harness. The source also does not establish whether the reported long-running behavior is available under ordinary customer operating conditions.
Skill to build now
Build a small, permissioned evaluation loop for security-related coding work this week. Use representative bugs or intentionally vulnerable test projects; require the agent to identify evidence, propose a patch, run tests, and leave a reviewable change set. Score both task completion and the human effort needed to verify the result. That gives a team a baseline for Carbon or any later model, instead of relying on aggregate benchmark claims.
Reader takeaway
Gemini 4 Carbon is a consequential frontier-model announcement for coding and cyberdefense, but not yet a broadly deployable product. Its practical value today is a prompt to prepare rigorous, repository-level evaluations for long-running and security-sensitive agent work.
Open question
When Google expands access, will Carbon's reported cyberdefense and reliability gains persist in external agent harnesses with real permissions, tools, and review workflows?