# Gemini 4 Carbon Gemini 4 Carbon is [[Google DeepMind]]'s second Gemini 4 model, announced on 28 October 2026, two weeks after [[Gemini 4 Argon]] became generally available. Google calls it "our strongest model with state-of-the-art cybersecurity and long horizon capabilities". It's built for long-running engineering and coding tasks, and for finding, exploiting and patching vulnerabilities. The headline came from [[Artificial Analysis]]: Carbon scores 59 on its Intelligence Index v5, one point above [[Claude Opus 5.5]] (58), which makes it the new number one. The catch is the same as with Argon at launch: almost nobody can use it yet. ## Key facts - **Access**: a subset of cyber defenders in Google's Fairwind Program first. Not in the Gemini app or the public API at announcement, and no date for a wider release. Argon took about two weeks to go from announcement to the API - **Pricing** (per Artificial Analysis): $4 per million input tokens and $20 per million output tokens, the same as Argon's standard price. Cached input is 95% off ($0.20 per million). Google offers 50% off at launch ($2 / $10), with no confirmed end date - **Context window**: 1M tokens - **Modalities**: text, image, video and speech in; text out - **Google's claims**: state-of-the-art on cybersecurity tasks, and secure code generated at "more than twice the success rate of Gemini 4 Argon on internal benchmarks". Secondary coverage cites 88.2% on SEC-bench Pro from Google's chart; I couldn't check it against the public leaderboard ## Artificial Analysis results Artificial Analysis tested Carbon at "high" reasoning (28 October 2026): | Metric | Gemini 4 Carbon | Comparison | |---|---|---| | Intelligence Index v5 | 59 (#1) | Claude Opus 5.5 (max) 58, [[GPT-6 Astra]] (max) 57, [[GPT-6.1 Sol]] (max) 55, Argon (high) 50 | | Coding Agent Index (in [[Antigravity CLI]]) | 59 | [[Claude Code]] with Opus 5.5 (max) 62 | | Terminal-Bench 4.0 | 64% (2nd) | behind Opus 5.5 (max) | | FrontierCode | 51% (2nd) | Opus 5.5 (max) ahead | | AA-Briefcase (agentic knowledge work, Elo) | 1609 | Argon 1483; Kimi K3.1 Pro 1678 and Muse Wonder 1659 ahead | | AA-Omniscience v1.1 | 54 (#1) | Opus 5.5 (max) 48, Argon 46 | | Cost per Index task | $7.76 | Opus 5.5 (max) $10.04, GPT-6 Astra (max) $6.25, Argon $7.68 | | Output tokens per Index task | 123k | Opus 5.5 (max) 173k | It leads four of the nine Index evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. The number I find most interesting is hallucination again. On AA-Omniscience v1.1, Carbon answers 58% correctly (10 points more than Argon) and hallucinates 11% of the time. That's worse than Argon (4% on v1.1) but far better than Opus 5.5 and GPT-6 Astra, both at 50%. Google's Gemini 4 models still say "I don't know" much more often than the competition. Careful though: Hacker News commenters report plenty of hallucinations with Argon in the Gemini app, so the benchmark and the product experience don't fully match. ## Read the #1 with care - **The index changed the day before.** Artificial Analysis launched Intelligence Index v5 on 27 October 2026, one day before Carbon's announcement. v5 has four equal categories (knowledge work, coding, science, general), adds FrontierCode 2.0 and Terminal-Bench-Science (15% of the whole index), and drops GDPval-AA, SciCode and AA-LCR. 60% of it now relies on private test sets. Carbon leads Terminal-Bench-Science, so the new weighting helps it. On Hacker News, rbehrends pointed out that this one science benchmark moved a lot, for a skill most users don't need - **One point is not a lead.** 59 vs 58 is a tie in practice. On the coding agent index, Claude Code with Opus 5.5 is still ahead - **Conflict of interest.** Artificial Analysis also sells benchmarking services to labs, and most of the index is now private. That's good against contamination, but it also means you can't reproduce the numbers - **You can't test it.** A leaderboard entry for a model only a few cyber defenders can access doesn't help you pick a model today. Several HN commenters asked Artificial Analysis to leave unreleased models out ## My take Google is clearly back at the frontier, and the hallucination numbers are the real differentiator: if a model knows when it doesn't know, I can trust it with research and knowledge work. But "number one" here means "one point ahead on a benchmark that changed the day before, for a model you can't use". The pattern repeats from Argon (and HN noticed the timing, right after Alphabet's earnings): announce, top a leaderboard, ship weeks later. I'll wait for API access and independent long-horizon evals before I try it with my own agents. Until then, Opus 5.5 stays my default for coding, and Argon is the Gemini you can actually use. ## References - Artificial Analysis, "Gemini 4 Carbon is the new #1 in the Artificial Analysis Intelligence Index" (2026-10-28): https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1 - Artificial Analysis, Intelligence Index v5 (2026-10-27): https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5 - Google, "Introducing Gemini 4 Carbon" (2026-10-28): https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ - Hacker News discussion of the launch: https://news.ycombinator.com/item?id=50276792 - Hacker News discussion of the Artificial Analysis result: https://news.ycombinator.com/item?id=50288019 - Khaleej Times (Reuters), "Google launches Gemini 4 Carbon AI model with cybersecurity focus" (2026-10-28): https://www.khaleejtimes.com/business/google-launches-gemini-4-carbon-ai-model-with-cybersecurity-focus-2 - Vals AI model page: https://www.vals.ai/models/google_gemini-4-carbon ## Related - [[Gemini 4 Argon]] - [[Gemini]] - [[Google DeepMind]] - [[Google]] - [[Artificial Analysis]] - [[Claude Opus 5.5]] - [[GPT-6 Astra]] - [[GPT-6.1 Sol]] - [[Antigravity CLI]] - [[Claude Code]] - [[AI Hallucination]] - [[AI Safety]] - [[Goodhart's Law]] - [[AI Frontier Model]] - [[Large Language Models (LLMs)]]