Gemini 4 Carbon: Google's new top model, what it scores and when you can use it

Picture a small software agency that looks after a dozen client websites and apps. Most weeks, the hardest job is not building something new. It is tracking down a bug that only appears under load, or patching a security flaw in code nobody on the team wrote. Those are long, fiddly tasks that need patience across many files and many hours. Google's newest model is built for exactly that kind of work.

Google announced Gemini 4 Carbon on October 28, 2026, calling it its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. It builds on Gemini 4 Argon, released two weeks earlier. Google says its teams have seen Carbon run for days on complex engineering tasks, such as finding and patching subtle bugs across large codebases. Like Argon, it is going first to a small group of cyber defenders, with developers, businesses and consumers to follow. ¹

Who can use it, and when

Almost no one, for now. Starting this week, Google is giving Carbon to a subset of cyber defenders and will gradually add trusted partners through its Fairwind Program, which gives high-priority defenders such as governments, healthcare providers and telecoms early access to advanced models. Google says it is taking part in the U.S. government's voluntary pre-release testing process and will keep refining guardrails before releasing Carbon to developers, businesses and consumers as soon as possible. It has not given a date. ¹³

Argon's path is a useful guide. It went from cyber defenders only to general availability in two weeks. Carbon may move as quickly, but plan around what is available today, not the announcement.

The independent ranking: now number one

Artificial Analysis switched to version 5.0 of its Intelligence Index this week, a new mix of nine evaluations. Carbon scores 59 on it, the highest of any model, just ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. It is a nine-point jump over Argon's 50. Because version 5.0 is new, these scores are not comparable with the version 4.3 figures in our earlier Gemini articles. ¹²

Two results stand out in Artificial Analysis's data. First, honesty: Carbon's hallucination rate on AA-Omniscience is 11.4%, second only to Argon's 4.3% among the models we compare, and far below Claude Opus 5.5's 50.3% and GPT-6.1 Sol's 51.9%. Second, getting work done: it posts the top score among these models on AutomationBench, a test of end-to-end business workflows, at 77.8%, just ahead of Argon's 77.5%. On Terminal-Bench 4.0 for coding agents it scores 65.2%, behind only Kimi K3.1 Pro's 67.2% in our comparison. ²

What it will cost

Google says Carbon will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input 95% off. After the introductory period, the price rises to $4 and $20. Artificial Analysis already prices both Gemini 4 models at that standard rate, which puts Carbon at about $7.76 per index task, against $10.04 for Claude Opus 5.5 and $6.25 for GPT-6 Astra. At the introductory price, simple halving puts it near $3.88. ¹²

One cost driver to watch: Carbon is wordy. Artificial Analysis recorded about 123,000 output tokens per index task, slightly more than Argon and more than twice GPT-6 Astra's 52,000. Output tokens are the expensive ones, so judge it on cost per finished task, not the price per million. ²

Built for cybersecurity defense

Carbon's headline strength is security. Google reports state-of-the-art results across the defense lifecycle, evaluated without guardrails: 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for checking which ones can actually be exploited, and 71.0% on CWE-bench for writing fixes at the root cause. ¹

Google's internal results point the same way. It says Carbon more than doubles Argon's success rate at generating secure code, that its Chrome team found Carbon produced high-quality patches more than ten times better than another lab's large frontier model, and that Wiz, its cloud security company, sees Carbon leading on web penetration-testing benchmarks. These are Google's own and partner claims. ¹

For a small business, that matters indirectly at first. Once Carbon is available, security tools and IT providers will use it to find and fix weaknesses faster, including in the plugins, themes and libraries that small business websites depend on. If you run your own software, it is worth asking your developer or IT provider how they plan to use models like this to check your code.

More careful behavior, by Google's measures

Google published unusually specific numbers on how Carbon behaves compared with Argon, all from its own testing. Several address the things that make AI agents hard to trust: cutting corners, making things up and doing more, or less, than asked. ¹

Google also says Carbon is better at handling ambiguity: rather than guessing, it gathers the relevant sources and points out discrepancies before acting, and it lists what is still needed instead of papering over missing evidence. Those are the behaviors you want from a model that may run for hours or days without someone watching. ¹

What this means for your business

Until Carbon opens up, Gemini 4 Argon is the closest thing you can use today, and much of our earlier advice still applies: budget on the standard price, not the launch discount, and compare total cost per finished task on your own work.