A two-week leap in the Gemini 4 cycle
Google DeepMind’s Gemini 4 Carbon landed on October 28, 2026, only two weeks after Gemini 4 Argon, turning what might have looked like a conventional model upgrade into a statement about release velocity. Google describes Carbon as its “next frontier model” and its strongest model so far, with particular emphasis on cybersecurity and long-horizon capabilities . The timing matters: Argon had already been framed as a major Gemini 4 milestone, but Carbon is being positioned as a fast iteration that incorporates user feedback, internal engineering lessons and improved safeguards rather than a distant generational successor .
The working headline of this story is therefore the story itself: Google DeepMind releases Gemini 4 Carbon just two weeks after Argon with gains across the board. The key claim is not simply that Carbon exists, but that Google is using it to show a shorter frontier-model cadence. In its announcement, the company says Carbon builds on Gemini 4 Argon, launched two weeks earlier, and illustrates how researchers and models are “working hand in hand” to advance capabilities faster while strengthening safeguards at the same time .
That framing will resonate across the AI industry because model releases have increasingly become not just technical milestones but competitive signals. A two-week interval between prominent Gemini 4 releases suggests Google wants the market to see the Gemini family as a rapidly compounding system. Carbon also arrives with enough measurable claims to make the release more than a branding update: a top score on Artificial Analysis Intelligence Index v5, cybersecurity benchmark wins, a WebDev Arena jump, internal secure-code gains and new safety metrics .
What Google says Carbon improves
Google’s central performance claim is broad: Carbon shows significant gains across coding, agentic tasks and cybersecurity defense, while also improving in scientific tasks, full-stack web development and multimodal understanding . In numerical terms, Google says Carbon ranks No. 1 on the recently updated Artificial Analysis Intelligence Index v5 with a score of 59 . Artificial Analysis independently describes Gemini 4 Carbon as the new No. 1 on that index and says the score is driven by strong performance in complex reasoning, knowledge and scientific tasks .
The cybersecurity numbers are the most concrete part of the launch. Google says Carbon scores 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching . Those three tasks map onto a full defensive workflow: detect a vulnerability, validate whether it is exploitable, and produce a fix. Anadolu Agency’s report on the launch also highlighted the same trio of cybersecurity scores, noting that Google described the model as strong in coding, autonomous agent tasks and cybersecurity defense .
Carbon is not just a cyber model, however. Google says it was designed for sustained agentic workflows, including complex engineering jobs that may involve dead ends, large codebases and multi-step root-cause analysis . StreetInsider, summarizing the announcement, reported that internal teams used Carbon to run multi-day workflows for engineering tasks such as identifying and fixing bugs across large codebases . That long-horizon angle is central to the release: Google is trying to show a model that can keep working without taking shortcuts, losing context or fabricating progress.
Benchmark gains and the “across the board” question
“Gains across the board” should be read carefully. Google’s launch materials argue that Carbon improves broadly over Argon and across many measured capabilities, not that it wins every public benchmark in existence. Artificial Analysis says Carbon leads the Intelligence Index v5 overall, but it also notes that Carbon is not leading in all coding and knowledge-work areas . That nuance matters because frontier-model rankings increasingly depend on which benchmark, mode, toolchain and cost assumptions are used.
Still, the breadth of improvements is notable. Artificial Analysis says Carbon leads four of the nine evaluations inside its Intelligence Index v5: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA . It also places Carbon second only to Claude Opus 5.5 on the Artificial Analysis Coding Agent Index when run with Antigravity CLI, with Carbon scoring 59 compared with 62 for Claude Code using Opus 5.5 . That is a strong showing for Google in an area where Artificial Analysis says Google had historically underperformed .
The model’s WebDev Arena movement gives another signal. Google says Carbon gains 89 Elo points to reach No. 4 on WebDev Arena, with better output quality and artifacts in web development . 9to5Google highlighted that jump in its launch coverage, along with Google’s claim that Carbon was improved across coding, agentic work, scientific tasks, full-stack development and multimodal understanding . In other words, the release is being sold as a broad usability upgrade, not only a benchmark chase.
Cybersecurity as the first deployment lane
The most consequential product decision is that Carbon is not immediately becoming a general-purpose model for everyone. Google says it will start by making Carbon available this week to a subset of cyber defenders, then gradually expand access to trusted partners through the Fairwind Program . The company also says it is participating in the U.S. government’s voluntary process for pre-release model testing . That phased rollout is a clear signal that the model’s strongest capabilities are also the ones Google sees as requiring the most careful deployment.
Cybersecurity is a double-edged domain for frontier AI. The same model that can find vulnerabilities and patch code can also validate exploitability. Google explicitly frames Carbon around the defense lifecycle, and says its external cybersecurity evaluations were run without guardrails . By starting with trusted cyber defenders, Google is trying to collect operational feedback while keeping the earliest access away from a completely open user base .
The safety claims are unusually detailed. Google says Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, with attempts to cheat dropping to 0.4%, while also improving overall accuracy . It also says high-severity misalignment alerts in live internal coding environments fell by almost 80%, unsupported claims were cut by approximately 2x compared with Argon, overreach fell by 26%, and under-service fell by 29% . These figures will need independent scrutiny, but they show where Google wants the conversation to move: not just raw model power, but reliability under pressure.
Pricing and access: fast, but not fully open
Carbon will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at a 95% discount . After the introductory period, Google’s footnote says the price will move to $4 per million input tokens and $20 per million output tokens . 9to5Google and Anadolu Agency both reported the same introductory and post-introductory pricing structure .
Artificial Analysis uses standard pricing in its charts and says Carbon’s standard rate is $4 per million input tokens and $20 per million output tokens, with the 50% launch discount and 95% cache discount layered on top . It calculates Carbon’s cost per Intelligence Index task at $7.76, about 23% less than Claude Opus 5.5 under its standard-pricing comparison, while scoring nine points higher than Gemini 4 Argon on the same index . That cost-performance angle is strategically important because frontier models are increasingly evaluated not just by headline scores, but by dollars per successful task.
The access story is more conservative. Google says Carbon is “coming soon” to developers, businesses and consumers through product surfaces, but the first availability is limited to selected cyber defenders and trusted partners . Artificial Analysis likewise notes that Carbon is being rolled out to selected users and is not yet publicly available . For most developers, then, Carbon is a launch announcement and a preview of the next Gemini tier, not yet a model they can immediately build against at scale.
Why Carbon matters for Google’s AI position
Gemini 4 Carbon matters because it gives Google a new top-of-the-stack narrative at a moment when model leadership is contested across multiple axes: reasoning, coding agents, scientific work, safety, price and deployment discipline. A single benchmark lead would be useful marketing. But Carbon combines several signals: No. 1 on Artificial Analysis Intelligence Index v5, strong cybersecurity scores, a WebDev Arena jump, internal alignment improvements and a two-week gap from Argon .
That combination strengthens Google DeepMind’s position in the frontier race. The message is that Google can iterate quickly, fold in feedback from a recent model, and push gains into both general intelligence measures and specialized cyber-defense tasks. It also shows a release strategy that is not purely consumer-first. By routing Carbon first through cyber defenders and the Fairwind Program, Google is treating frontier deployment as a staged process rather than a switch flipped for all users .
There are caveats. Some of Carbon’s most striking claims remain Google-reported internal measurements, not independently reproduced public evaluations. The Chrome patching comparison, the multi-day workflow behavior, the misalignment-alert reduction and the secure-code uplift all need outside validation . Artificial Analysis offers a stronger independent view on the Intelligence Index and coding-agent comparisons, but even there, rankings depend on specific test suites, reasoning settings and tool environments .
Still, Carbon is a meaningful release. It indicates that Argon was not the endpoint of Gemini 4 but the base for a faster iteration loop. If Google can deliver broader access without losing the safety and reliability improvements it claims, Gemini 4 Carbon could become the model that defines the next phase of Google’s AI strategy: faster cycles, specialized defensive deployment, stronger agentic coding and a more aggressive claim to frontier leadership.