{"article":{"slug":"google-gemini-4-carbon-limited-release","kind":"model","source":{"title":"Gemini 4 Carbon: our next frontier model","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/","date":"2026-10-28"},"headline":"Gemini 4 Carbon leads a code-patching chart by 3 points, but few customers can use it","whatChanged":"Google announced Gemini 4 Carbon on October 28 and opened it only to a subset of cyber defenders. Access will widen gradually to trusted partners through the Fairwind Program. Google gave no date for developers, businesses or consumers and says only \"as soon as possible\". It says it is taking part in the U.S. government's voluntary pre-release testing process. Pricing is $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper, the same introductory rate as Gemini 4 Argon. A footnote sets the price after the introductory period at $4 and $20. The post gives no end date for the introduction.\n\nGoogle says Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 at 59, against 58 for Claude Opus 5.5 and 57 for GPT-6 Astra. Its security results are SEC-bench Pro 88.2%, ExploitGym 54.8% and CWE-bench 71.0%, which Google describes as \"evaluated without guardrails\". Google's methodology PDF says it ran SEC-bench Pro and ExploitGym itself in its Antigravity harness and took rivals' numbers from their system cards and technical reports. Collinear AI ran CWE-bench across all models, each in its own harness.","whyItMatters":"Few readers can use Carbon today. Access is limited to cyber defenders and Google gives no wider date, so a purchase decision now would rest on a model you cannot test. The parts that apply to everyone are the price path and the evidence. The launch price doubles after an introductory period of unstated length, which matters for any budget built on the $2 and $10 rates.\n\nOn evidence, CWE-bench is the one security result run by a third party: Carbon 71.0 against 67 to 68 for four rivals, each in its own tooling. Google's other security numbers are its own runs.","issuer":"Google","glance":"Google announced Gemini 4 Carbon October 28 for a subset of cyber defenders only, at an introductory $2 in and $10 out.","keyFact":"$2/$10 per 1M tokens, then $4/$20","action":"Plan any Gemini 4 budget at $4 and $20 per million tokens, not the introductory rates.","visuals":[{"url":"https://storage.googleapis.com/gweb-uniblog-publish-prod/images/gemini_4_carbon_cwe-bench.original.png","alt":"Bar chart of CWE-bench Pass@1 for 11 model and harness pairs. Gemini 4 Carbon in Antigravity leads at 71.0%; Grok 4.7, Gemini 4 Argon and GPT-6 Astra score 68%, Claude Opus 5.5 67%, and GPT-6 Sol is last at 52%.","caption":"Google's CWE-bench chart, October 28. Pass@1 with ties broken by Pass@4, higher is better. Carbon 71.0, Grok 4.7 68, Argon 68, GPT-6 Astra 68, Opus 5.5 67, Fable 5.1 58, GPT-6 Sol 52. Collinear AI ran the benchmark, with each model in its own harness (Carbon and Argon in Antigravity, Astra and Sol in Codex, Opus and Fable in Claude Code, Grok in opencode), so tooling differs. Google published the chart and highlights its own bars.","credit":"Google DeepMind","source":{"title":"Gemini 4 Carbon: our next frontier model","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/","date":"2026-10-28"}},{"url":"https://storage.googleapis.com/gweb-uniblog-publish-prod/images/gemini_4_carbon__evals-table.original.png","alt":"Table of 17 benchmark scores for Gemini 4 Carbon, Gemini 4 Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. Carbon leads 15 rows; GPT-6 Astra leads FrontierSWE 2 and Claude Opus 5.5 leads the OSWorld 2.1 offline subset.","caption":"Google's benchmark table, October 28. Carbon: AA Index v5.0 59, Terminal-Bench 4.0 65.2%, Agents' Last Exam v1.1 77.4%, ExploitGym 54.8%, SEC-bench Pro 88.2%. It trails GPT-6 Astra on FrontierSWE 2 (61.2% to 65.5%) and Opus 5.5 on the OSWorld 2.1 offline subset (81.5% to 87.2%). Gemini ran at its highest thinking setting; rivals' numbers mostly come from their own reports or public leaderboards. Rows use different units and should not be combined.","credit":"Google DeepMind","source":{"title":"Gemini 4 Carbon: our next frontier model","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/","date":"2026-10-28"}}],"attachments":[{"title":"Gemini 4 Carbon: approach, methodology and results","url":"https://deepmind.google/models/evals-methodology/gemini-4-carbon","description":"Google DeepMind's five-page methodology PDF: which scores Google ran itself, which come from leaderboards, the harnesses and attempt counts, and the thinking settings used.","source":{"title":"Gemini 4 Carbon: our next frontier model","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/","date":"2026-10-28"}}],"companyLabel":"Public edition"},"issue":{"id":"2026-11-02","publishedLabel":"Week of November 2, 2026","isLatest":false},"position":{"index":2,"total":4},"next":{"slug":"mistral-large-4-open-weights","headline":"Mistral releases Large 4 weights under a modified MIT license, with a catch for resellers"},"session":null}