Announcing Gemini 4 Carbon

Google opened Gemini 4 Carbon to everyone on November 10, 2026, two weeks after unveiling it — and only two weeks after Gemini 4 Argon shipped. It is Google's strongest model to date, built around two specialties: cybersecurity and long-horizon agentic work that can run for days.

What Carbon does

Carbon takes text, images, video, and speech as input and outputs text through a unified 1 million-token window shared between input and output. Thinking depth is adjustable across low, medium (default), and high levels, and generation runs at about 94.6 tokens per second.

The model debuts at #1 on the Artificial Analysis Intelligence Index v5.0 with 59 points, and Google claims first place in 15 of 17 tracked benchmarks. Independent testing largely confirms the lead, with two caveats: Carbon trails Claude Opus 5.5 on the Coding Agent Index (59 vs 62, though at roughly 23% lower cost per task) and sits second on the Vals Index behind Sonnet 5.5 — while still holding #1 on Legal Research Bench. It is also the least hallucination-prone model in the AA-Omniscience test.

Cybersecurity, the headline feature

Carbon posts state-of-the-art results across the full defense cycle: 88.2% on SEC-bench Pro (vulnerability discovery), 54.8% on ExploitGym (validation and exploitation), and 71% on CWE-bench (root-cause fixes). Google reports more than a 2× improvement in secure-code generation over Argon, the Chrome team says it produces quality patches more than 10× better than another lab's frontier model, and security firm Wiz used it to find previously unknown vulnerabilities in two major SaaS providers in under two hours.

Safety numbers

• Prompt injection success rate: 0.0% on Gray Swan IPI (Argon was at 0.7%)
• Automated cyberattack success cut to 2.17%, down from 9.86%; Trajectory Labs found no new universal jailbreaks in 132 hours and over 13,000 attempts
• Honesty: 92% of tasks passed on an internal 375-scenario agentic benchmark (Argon: 70%)
• Plan execution: 26% less overstepping of boundaries and 29% less under-delivery

Pricing and availability

Until November 30, 2026 the Gemini API charges an introductory $2 per million input tokens and $10 per million output, with cached input discounted 95%. From December 1 the rates become $4/$20 — about half of Claude Fable 5.2's $8/$40 and well under GPT-6 Astra's $10/$50.

Access runs through the paid Gemini API tier in AI Studio and Android Studio. Google AI Ultra subscribers also get Carbon inside Antigravity — desktop, CLI, and IDE extensions — including multi-agent /swarm commands. In pre-release evaluations it matched the top model on Hebbia's end-to-end finance test and ranked first of 23 configurations in a Balyasny Asset Management benchmark; in one Antigravity demo it built an interactive 3D textile-structure visualizer from an academic paper off a minimal prompt.

Where to try Gemini 4 Carbon in Russia — no VPN needed

• ru.vega.chat — AI model API service
• officevega.com — the first office suite with AI
• vega.codes — vibe coding and AI agents