Gemini 4 Carbon Leads Cyber Benchmarks, But Access Is Gated

Google DeepMind released Gemini 4 Carbon on October 28, 2026, calling it the company's strongest model and a successor to Gemini 4 Argon, which launched two weeks earlier. Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect at Google, positioned Carbon as built for "state-of-the-art cybersecurity and long horizon capabilities." Google has not disclosed Carbon's parameter count or context window, and initial access is limited to a vetted subset of cyber defenders this week, with broader rollout later through the Fairwind Program.

Google's published comparison puts Carbon first on Artificial Analysis Intelligence Index v5.0 with 59, ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57 and Argon at 50, and shows Carbon leading 15 of 17 published benchmarks. Carbon and Argon ran at Google's highest thinking setting; rival models ran at max available reasoning, with most scores self-reported under Google's evaluation methodology.

Coding and long-horizon performance

Google reports Carbon gains 89 Elo points on WebDev Arena to reach #4 at 1,768. On Terminal-Bench 4.0, Google reports Carbon scores 65.2%, ahead of Claude Opus 5.5 at 59.6% and GPT-6 Astra at 59.1%; GPT-6 Astra leads FrontierSWE at 65.5% to Carbon's 61.2%. The widest published gap is GraphWalks at 256K to 1M tokens, where Google reports 96.2% BFS F1 versus 84.2% for Argon and 71.8% for GPT-6 Astra. Carbon also ends Vending-Bench 2 with the highest bank balance in the comparison, $16,484, while OSWorld-2.1 is the notable miss at 81.5% against Claude Opus 5.5's 87.2%. Google says internal teams saw Carbon work for days on tasks such as root-causing and patching low-level bugs.

Cybersecurity benchmarks

Google frames Carbon around finding vulnerabilities, validating exploitability, then patching. All three cybersecurity benchmarks were evaluated without guardrails. Google reports 88.2% on SEC-bench Pro, ahead of GPT-6 Astra at 85.4% and Argon at 83.4%; 54.8% on ExploitGym, ahead of GPT-6 Astra at 42.4% and Claude Opus 5.5 at 34.5%; and 71.0% on Collinear's CWE-bench, ahead of Argon and GPT-6 Astra at 68.0%. Google self-computed the SEC-bench Pro and ExploitGym scores with its Antigravity harness, and ExploitGym ran with mitigations off. The Chrome team reports Carbon was more than 10x better than another lab's large frontier model at producing high-quality patches, and Wiz reports it leads its web penetration testing benchmarks.

Pricing and access

Carbon launches at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off the input rate. Google's footnote sets the post-introductory price at $4 input and $20 output. Access is deliberately staged. Google is participating in the U.S. government's voluntary pre-release model testing; access this week goes to a subset of cyber defenders, then to vetted partners through the Fairwind Program, covering governments, critical infrastructure operators and core tech platforms. Developers, businesses and consumers are deferred with no date. This is part of a broader gated capability verification pattern for dual-use frontier launches.

AI Mastery reference pricing as of 2026-10-22, except Gemini 4 Carbon which is sourced from Google's launch post. Post-intro Carbon pricing from Google's footnote.

Safety changes over Argon

Google reports Terminal-Bench 4.0 successful cheating fell to 0.0% with attempts at 0.4% while accuracy improved, high-severity misalignment alerts in live internal coding fell almost 80%, and unsupported claims fell about 2x. Overreach dropped 26% and under-service 29%. On Gray Swan's Indirect Prompt Injection benchmark, Carbon's attack success rate is 0.0%, down from Argon's 0.7%. All are Google internal or self-run results.

AI Mastery analysis

At $2/$10, Carbon is priced identically to Claude Sonnet 5, a mid-tier model, while Google claims frontier-level benchmark results. The post-introductory $4/$20 rate, confirmed in Google's footnote, remains below GPT-6 Astra and Claude Fable 5.1 at $10/$50 in AI Mastery's pricing data. Teams should budget against $4/$20 because no intro end date is published and long-horizon agentic workloads consume substantial tokens.

The security scores also need context. Google self-scored SEC-bench Pro and ExploitGym with mitigations disabled. The 54.8% ExploitGym figure is therefore an unconstrained capability result, not a production posture; the infrastructure isolation question remains live for defenders running Carbon in monitored pipelines. Finally, the Fairwind-only rollout creates a two-tier market: vetted partners get an early integration window, while everyone else waits for a general API date Google has not announced.

Sources

• Primary: Introducing Gemini 4 Carbon
• Reported by: MarkTechPost: Google DeepMind Introduces Gemini 4 Carbon, Its Top Model for Cyber Defense and Coding
• Competitor prices: AI Mastery token calculator, figures as of 2026-10-22