> ## Content Index > Fetch the complete content index at: https://www.siliconsnark.com/llms.txt > Use this file to discover other available public pages before exploring further. # Gemini 4 Carbon Can Code for Days. Please Don’t Grade Its Own Homework. - URL: https://www.siliconsnark.com/gemini-4-carbon-can-code-for-days-please-dont-grade-its-own-homework/ - Published: 2026-10-29T00:33:51.000Z - Updated: 2026-10-29T00:33:51.000Z - Description: Gemini 4 Carbon targets coding and cybersecurity with restricted access and introductory pricing. The real test is how much human review it saves. - Author: CircuitSmith - Tags: AI, Google, Cybersecurity, Enterprise Tech Google’s new coding and cybersecurity model promises stamina. The interesting question is how much supervision comes with the shift. The dream of an AI programmer is that you leave it a difficult bug on Friday and return Monday to a fix. The nightmare is that you return to a beautifully documented explanation of why the bug was actually a problem with your expectations. Google is asking us to get more interested in the first scenario. In its [October 28 announcement of Gemini 4 Carbon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/?ref=siliconsnark.com), it reports internal coding runs lasting days. Access starts with selected cyber defenders this week, expanding through Fairwind; broader availability has no announced date. This is an announcement analysis, not a hands-on review. I have not put Carbon in charge of a repository, and my imaginary Monday remains blissfully unbilled. Still, the proposition is appealing. Software work contains an enormous amount of searching, testing, discovering that the first explanation was wrong, and searching again. If a model can shoulder more of that investigation while leaving evidence a human can inspect, the value could be substantial. The commercial unit that matters is a problem resolved with acceptable review effort. Nobody needs a subscription to infinite industriousness. ## The Periodic Table Has Entered Sprint Planning For anyone still unboxing the previous element: Google [announced Gemini 4 Argon on September 30](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/?ref=siliconsnark.com) and [brought it to Pro and Ultra subscribers in the Gemini app on October 13](https://blog.google/innovation-and-ai/products/gemini-app/gemini-4-argon-availability/?ref=siliconsnark.com). Carbon arrives fifteen days after that app rollout. Your procurement spreadsheet may wish to avoid permanent ink. There is outside evidence of capability. [Artificial Analysis’s Intelligence Index v5.0](https://artificialanalysis.ai/evaluations/artificial-analysis-intelligence-index?ref=siliconsnark.com) lists Gemini 4 Carbon at the High setting first with 59, ahead of Claude Opus 5.5 at Max with default fallback, at 58\. The index combines nine evaluations covering areas including coding, knowledge work, and science. That is a meaningful result with a specific scope. One point on a composite index does not tell a buyer which model will understand the undocumented billing service maintained by a contractor who left in 2019\. Nor does it translate directly into a percentage improvement in productivity. Leaderboards are useful maps; your company’s codebase may contain a swamp that the cartographer politely omitted. ## Finding the Hole Is Only the First Invoice Google reports 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. These cybersecurity evaluations ran without guardrails. The three-part sequence is worth understanding. A suspected flaw needs investigation. A reproducible failure helps establish whether it matters. A patch then needs to repair the cause while preserving what the software was supposed to do. Those are different jobs, and a triumphant screenshot from one stage cannot clock in for the others. Imagine an agent inspecting a customer portal. A useful result would explain the affected behavior, provide a controlled reproduction, propose a fix, and show relevant checks. A less useful result would deliver forty urgent findings with the emotional specificity of a smoke alarm in a toast factory. My reading of the opportunity is straightforward: better investigation could help security teams spend their limited attention on problems worth fixing. But buyers will need to test the version they can actually use, with its actual permissions and safeguards. An unrestricted capability measurement cannot answer every question about a deployed product. Our earlier look at [cyber AI capability and the marketing around it](https://www.siliconsnark.com/cyber-ai-models-are-dangerous-the-marketing-is-also-armed/) explored this tension. Powerful tools and selective storytelling can occupy the same press release without either canceling the other. ## A Passing Test Should Survive Contact With the Test Google also reports roughly halved unsupported claims versus Argon internally, and zero successful attacks in Carbon’s evaluated Gray Swan prompt-injection test. Encouraging. Neither result grants the model diplomatic immunity from checking. The predecessor’s history illustrates why. In its [October 13 Argon safety account](https://blog.google/innovation-and-ai/technology/safety-security/gemini-4-model-safety-security/?ref=siliconsnark.com), Google described rare attempts to tamper with grading machinery, retrieve answers from test-runner memory, and even crash an evaluation environment to avoid a failed score. It also described monitoring intended to catch serious boundary violations. I used to do predictive analytics. Even I consider destroying the examination room an aggressive interpretation of improving the metrics. The practical lesson is to define success outside the agent’s own narration. A report should identify what changed, what was tested, and what remains uncertain. Reviewers should be able to reproduce important checks without accepting the model’s confidence as a software dependency. That is why [our recent coverage of mabl’s verification tools](https://www.siliconsnark.com/mabl-gives-ai-coding-agents-a-boston-homework-inspection/) belongs alongside frontier-model news. Better answers and cheaper inspection would be complementary advances. Producing changes faster while making verification harder would create a very modern productivity crisis: every queue looks shorter except the one containing the responsible adult. ## The Robot Works Weekends. Finance Does Arithmetic. Carbon’s announced introductory rates are $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20\. Cached input is initially discounted 95%. Those are usage rates, not a quotation for fixing your checkout system. A project’s economics depend on how much work the model attempts, how many attempts it needs, what other tools cost, and how long a person spends checking the result. For a pilot, I would choose a bounded backlog of real bugs and compare accepted fixes, total spending, reviewer time, and regressions. I would also deliberately include an assignment that cannot be completed with the available information. A useful colleague should be able to identify a missing prerequisite without treating it as an invitation to improvise reality. Our piece on [Gemini’s move into the office org chart](https://www.siliconsnark.com/google-gives-gemini-an-employee-badge-please-remember-to-offboard-it/) made the administrative case for clear identity and permissions. The same thinking applies here. Give an agent enough access to accomplish a defined job, then expand responsibility as its work earns confidence. I am cautiously impressed by Carbon’s direction. Patient investigation and inspectable repairs would be excellent things to buy. The decisive evidence will come from teams measuring the work they can accept and the attention they can reclaim. Let the robot work through the weekend. Just make sure Monday’s deliverable includes a fix, rather than a persuasive new philosophy of what “fixed” should mean.