Gemini 4 Carbon Reaches the Top, but Developers Still Have to Wait

Google released Gemini 4 Carbon on October 28. It was only two weeks after the release of Gemini 4 Argon. The official positioning is straightforward: the strongest model, with a focus on cybersecurity and long-horizon tasks.

Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59. Google says it has clear improvements in coding, agentic tasks, and cybersecurity defense-related capabilities. Pricing follows Argon's introductory pricing: $2 per million input tokens, $10 per million output tokens, and a 95% discount on cached input. After the introductory period, it becomes $4 input and $20 output.

But the release approach is more worth watching than the scores. Carbon is first available to a subset of cybersecurity defenders, then gradually to trusted partners through the Fairwind Program. Developers, enterprises, and consumers have to wait for later product releases. Google says it is following the U.S. government's voluntary pre-release testing process.

Cybersecurity Is the Main Event

The official external benchmarks are: SEC-bench Pro 88.2%, ExploitGym 54.8%, and CWE-bench 71.0%. These three correspond respectively to finding vulnerabilities, verifying whether vulnerabilities can be exploited, and fixing confirmed issues. Google also mentions that the Chrome team reported Carbon is more than 10 times better than a large frontier model from another lab at generating high-quality patches; Wiz observed that it leads on a web penetration testing benchmark.

For security teams, this is a candidate for a toolchain that can directly run vulnerability discovery, exploitability validation, and patch generation. For ordinary developers, right now there is only a blog link.

Coding and Long-Horizon Tasks

Carbon continues to push agentic coding. On WebDev Arena it gained 89 Elo points to reach 1768, ranking 4th. Google's internal teams say Carbon can run continuously for several days, handling hard-to-pin-down bugs and low-level system issues in large codebases. This description is more specific than "can write code"; long-horizon tasks are the point it wants to sell.

Alignment and Safety

Google says Carbon's successful cheating rate on Terminal-Bench 4.0 dropped to 0.0%, attempted cheating dropped to 0.4%, and overall accuracy still improved. Internal high-severity misalignment alerts dropped by nearly 80%. Unsupported claims are about half as many as with Argon. Overreach dropped 26%, and under-service dropped 29%. On Gray Swan's Indirect Prompt Injection Benchmark, Carbon's attack success rate was 0.0%.

These numbers all come from Google's official blog and internal tests. External reproduction will have to wait until the model is more widely available.

The Replies Are More Real Than the Launch Post

Some users in the replies to Logan Kilpatrick's tweet said Carbon beats Opus 5.5 on almost every benchmark, and posted a video. This claim comes from the replies, not from an official Google conclusion.

Some also asked about the naming: since it is priced the same as Argon, why not call it 4.1 Argon? Another user posted an image complaining that the chart used low-brightness colors for competitors.

Some directly asked for access, and some said it should wait until cybersecurity defenders have used it before a public release. Model release frequency has been compressed to once every two weeks, and the names are starting to be harder to remember than the capabilities.

The scores are loud, and the barriers are very real. Carbon now looks more like a frontier tool for cybersecurity defenders than an off-the-shelf product for all developers. Official blog: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ . Logan Kilpatrick's tweet: https://x.com/OfficialLoganK/status/2115489284404940898 .