Gemini 4 Carbon: Cyber Benchmarks, Pricing and Access

Gemini 4 Carbon is Google's strongest model to date, announced on 2026-10-28. Google says it is built for sustained agent workflows and cybersecurity. Most numbers come from Google's own announcement, and you cannot call the model yet: access starts with a small group of cyber defenders.

Key Takeaways

• Google reports Carbon at #1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59. Artificial Analysis lists the same score; the other figures are Google-reported or come from a third-party leaderboard Google cites.
• Google reports 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym and 71.0% on CWE-bench (patching). On WebDev Arena it scores 1,768 Elo, ranked #4.
• Introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input 95% off. The post-introductory rate is $4 and $20.
• Access is limited to selected cyber defenders and trusted partners through the Fairwind Program. Google gives no date for general availability.

What is Gemini 4 Carbon?

Google's announcement calls Carbon "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." It builds on Gemini 4 Argon, which launched two weeks earlier, and keeps Argon's safety and security stack. Google says it added further alignment training for complex or high-pressure tasks.

Google lists four focus areas:

• Long-running agents: built for sustained agentic workflows. Google says internal teams run it "for days to persist through dead ends."
• Coding: long-horizon agentic coding, plus improved full-stack web development.
• Cyber defense: finding vulnerabilities and helping generate patches. The Chrome team reports more than 10x better patch generation than "another lab's large frontier model." That is a partner claim, not an independent test.
• Prompt-injection resistance: 0.0% attack success on Gray Swan's Indirect Prompt Injection (IPI) benchmark, which Google describes as "resisting every evaluated attack attempt." Argon scored 0.7% on the same benchmark. Google's post does not say who ran the evaluation.

What do Carbon's benchmark scores show?

Artificial Analysis lists the same score of 59, against a median of 38 for comparable models. Google says its external cyber benchmarks were "evaluated without guardrails," and several other figures are internal.

WebDev Arena is the one place Carbon does not lead. Google reports a gain of 89 Elo points, but the model ranks fourth, so the web-development gain is real but not a lead.

Google also reports internal gains over Argon:

• More than 2x uplift in secure-code success rate.
• Unsupported claims cut by about 2x.
• High-severity misalignment alerts down almost 80% in live internal coding.

On Terminal-Bench 4.0, Google says successful cheating dropped to 0.0%, with cheating attempts at 0.4%.

How much does Carbon cost?

Google's introductory API pricing is $2 per million input tokens and $10 per million output tokens, the same introductory rate Google lists for Argon. Cached input tokens get a 95% discount. Google's announcement lists the later rate as $4 input and $20 output per million tokens, and gives no end date for the introductory period. Artificial Analysis already prices the model at the later $4 and $20 rate.

For long agent runs that resend the same repository context on every step, the 95% cached-input discount applies to that repeated input, so it can cut the input bill far more than the base rate does.

Who can use Carbon today?

Only a limited group. Google says that, starting the week of the announcement, a subset of cyber defenders gets access, then gradually adding trusted partners through the Fairwind Program. Google says it is still gathering feedback and iterating on guardrails.

General availability has no date. Google says it wants broader access for developers, businesses and consumers "as soon as possible." Google's announcement gives no API model ID. Artificial Analysis's page is inconsistent on API availability, with its FAQ and charts disagreeing, so treat Carbon as not callable yet. Google also says it is taking part in the U.S. government's voluntary pre-release model testing process.

Is Carbon worth waiting for?

Carbon is worth tracking if you build coding agents or security tooling. Its independent score is strong, and its cyber and prompt-injection numbers are specific enough to test once you get access. Until then, build on a model you can call today, keep model IDs in config, and plan to rerun your own evals when Carbon opens up.