Google: Gemini 4 Carbon

Google made Gemini 4 Carbon generally available on November 10, 2026 to developers, enterprises, scientists, and consumers, following its earlier limited release to cybersecurity defenders via the Fairwind Program. The announcement positions Argon as the default choice for most software engineering and knowledge work, with Carbon aimed at harder multi-step coding, debugging, and open-ended scientific research using more tokens and extended verification loops. Google said Carbon ranks first on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench at more than 2X lower cost per task than the next best external model. The model is available via API in Google AI Studio and Gemini for business for all paying developers and enterprises at the same introductory rates as Argon, and Google is previewing /swarm in Google Antigravity for multi-agent coordination.

Cybersecurity News covered Google's October 28, 2026 unveiling of Gemini 4 Carbon, noting it builds on Gemini 4 Argon and targets complex security and coding work that can take days, with access beginning only with selected cyber defenders rather than a public launch. The report says Google's internal teams saw the model work through dead ends while tracing subtle bugs across large codebases and low-level systems before producing root-cause patches. The article repeats Google's reported figures of 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench, plus a lead on the Artificial Analysis Intelligence Index v5.0 with a score of 59, and adds an important caveat that Google evaluated these capabilities without guardrails, which matters when comparing results to deployed services.

Google announced Gemini 4 Carbon on October 28, 2026, describing it as its strongest model with state-of-the-art cybersecurity and long-horizon capabilities, building on Gemini 4 Argon which launched two weeks earlier. The post is attributed to Koray Kavukcuoglu, SVP of Google DeepMind and Google's chief AI architect. Google reported Carbon ranking first on the updated Artificial Analysis Intelligence Index v5.0 at 59, with vendor-reported scores of 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. Introductory pricing was set at $2 per million input tokens and $10 per million output tokens, with cached input discounted 95%, and rollout started with selected cyber defenders through the Fairwind Program.

OpenRouter lists Google Gemini 4 Carbon as a frontier Gemini 4 model for coding, agentic and complex reasoning work, released November 10, 2026. The listing states it accepts text, image, audio, video and file input, carries a 1.0M context window, and supports long decoding of up to 1M output tokens through continuation, with thinking depth set by low, medium and high reasoning effort levels. The page also describes OpenRouter's own routing modes for the model, including Balanced, Nitro and Floor, which select among hosted Google Vertex and Google AI Studio endpoints. It reports introductory rates of $2 per million input tokens and $10 per million output tokens with list pricing at $4 and $20, alongside per-endpoint latency, throughput and uptime measurements that vary by hosting tier.

Google announced Gemini 4 Carbon on October 28, 2026, describing it as its strongest model to date, with emphasis on cybersecurity and long-running coding work. The initial release went to a subset of cyber defenders, with additional trusted partners added through the Fairwind Program, and no general availability date was set at announcement. Artificial Analysis reportedly ranked it first on its Intelligence Index v5 with a score of 59. NeuralTrust reports introductory pricing of $2 per million input tokens and $10 per million output tokens, moving to $4 and $20 afterward, with cached input priced about 95% lower. It notes the model trails Claude Opus 5.5 on the Coding Agent Index, 59 versus 62, and that Google's cyber results — 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench — were measured without guardrails and have not been independently replicated.