Gemini 4 Carbon (Vertex AI)

Google's October 28, 2026 Gemini 4 Carbon introduction describes Carbon as its strongest frontier model for cybersecurity and long-horizon work, with an Artificial Analysis Intelligence Index v5 score of 59 and cyber-defense results including SEC-bench Pro at 88.2%, ExploitGym at 54.8%, and CWE-bench at 71.0%. Pricing was introduced at $2 per million input tokens and $10 per million output tokens, with cached input discounted 95%, before moving to $4/$20 later. A November 10, 2026 availability check finds Google's public Cloud documentation still listing Gemini 4 Argon rather than Carbon as the callable Gemini 4 family on Vertex AI, so a project routing traffic through Vertex AI likely still calls Argon. Carbon is confirmed callable on the Gemini API developer-key route, since Simon Willison's llm-gemini 0.36 ships a gemini-4-carbon model id, while Vertex and Fairwind enterprise access was not verified. Access expands from a subset of cyber defenders through the Fairwind Program before developers and consumers.

Security-focused coverage of Gemini 4 Carbon describes the model as built to find software flaws, test whether attackers could exploit them, and produce fixes for complex security and coding work that can take days. Google reported internal teams observing Carbon push through dead ends while tracing subtle bugs across large codebases and low-level systems, then patching root causes. The article highlights an important measurement caveat: Google evaluated these capabilities without guardrails, which matters when comparing the benchmark results with deployed services. Reported scores match the vendor's claims of 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench, plus a leading Artificial Analysis Intelligence Index v5.0 score of 59.

TechTimes reports that Gemini 4 Carbon launched at the top of the Artificial Analysis Intelligence Index v5.0 with a score of 59, making it the first model to hold that position under the revised v5.0 methodology. The announcement was authored by Koray Kavukcuoglu, who framed Carbon as a direct successor to Gemini 4 Argon, released two weeks earlier. The article outlines the phased rollout, beginning with a subset of cyber defenders through the Fairwind Program and widening to trusted partners before developers, businesses, and consumers get access via product surfaces. Introductory pricing is $2 per million input tokens and $10 per million output tokens with cached inputs at 95% off, rising to $4 and $20 after the introductory period.

Nerdschalk tracked Gemini 4 Carbon's release status, reporting that selected cyber defenders received access first, followed by additional trusted Fairwind Program partners, while public access to developers, businesses, and consumers has no announced date. A paid API key's October 28 model list included Gemini 4 Argon but no Carbon model, and Argon remains available. The piece also records the pricing structure, with introductory rates of $2 per million input tokens and $10 per million output tokens rising to $4 and $20 afterward, plus a WebDev Arena score of 1769 ranking fourth after a gain. It notes reported improvements in scientific tasks, multimodal understanding, full-stack web development, and generated web artifacts over Argon.

Google DeepMind announced Gemini 4 Carbon on October 28, 2026, describing it as its strongest frontier model with state-of-the-art cybersecurity and long-horizon capabilities. It ranks first on the updated Artificial Analysis Intelligence Index v5.0 with a score of 59, and the post is attributed to Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google. Carbon builds on Gemini 4 Argon, launched two weeks earlier, and reportedly leads SEC-bench Pro at 88.2% for vulnerability finding, ExploitGym at 54.8% for exploitation, and CWE-bench at 71.0% for patching. Introductory pricing matches Argon at $2 per million input tokens and $10 per million output tokens, with cached input at 95% off, while access starts with cyber defenders through the Fairwind Program.

Google DeepMind published a Gemini 4 Carbon model evaluation document covering agentic coding, computer use, knowledge work, science and math, multimodal capabilities, long context, cybersecurity, and the Artificial Analysis Intelligence Index. All Gemini 4 Carbon and Gemini 4 Argon scores are pass@1 unless otherwise noted, run with the Gemini API at the highest thinking settings. The evaluation explains its sourcing and setup in detail: SWE-sweep results for Carbon were self-computed with a mini-swe-agent harness, while FrontierSWE 2 numbers come from Proximal's public leaderboard and Terminal-Bench 4.0 figures from Artificial Analysis. Agents' Last Exam v1.1 results were self-computed on 122 tasks with a 5-hour window and safety filters enabled, noting non-Gemini model numbers are providers' self-reported.