Gemini 4 Carbon Took First Place, but Developers Still Can't Use It Yet

Google has released Gemini 4 Carbon, calling it the strongest frontier model currently available and saying it ranks first on the Intelligence Index v5.0 from third-party evaluation firm Artificial Analysis with a score of 59. The official positioning is clear: cybersecurity defense and long-horizon tasks.

The rollout cadence is worth noting. This week, Carbon is first being made available to a small group of cybersecurity defenders through Google DeepMind's Fairwind Program, and will later gradually expand to trusted partners. Developers, enterprises, and ordinary users can only wait for “soon.” Someone in the comments asked directly: will broader access take weeks or months? There is no answer.

Capabilities First

The numbers listed in Google's blog:

• SEC-bench Pro 88.2%, finding real vulnerabilities
• ExploitGym 54.8%, verifying exploitability of vulnerabilities
• Collinear's CWE-bench 71.0%, fixing root causes
• On internal benchmarks, the success rate for generating secure code is more than twice that of Gemini 4 Argon
• The Chrome team reports that Carbon's performance in generating high-quality patches is more than 10 times better than a large frontier model from another lab
• On WebDev Arena, it gained 89 elo to reach 1768, ranking 4th

These benchmarks each cover a different part of the defense workflow: finding vulnerabilities, validating risk, and patching. Google says Carbon has run multi-step engineering tasks lasting several days in internal coding environments, such as locating and fixing subtle bugs in large codebases.

Safety Alignment

Carbon is built on Argon's multi-layered safety stack. Google highlights several changes:

• On Terminal-Bench 4.0, cheating success was 0.0%, cheating attempts dropped to 0.4%, and overall accuracy still improved
• Internal high-severity misalignment alerts fell by nearly 80%
• Unsupported claims were roughly halved compared with Argon
• Over-intervention fell 26%, and under-service fell 29%
• On the Gray Swan indirect prompt injection benchmark, 0.0%, versus 0.7% for Argon

The 0.0% figure is striking. A benchmark is just one evaluation and cannot be treated as absolute security, but it is still meaningful compared with Argon's 0.7%.

Pricing and Access

Carbon's introductory pricing is the same as Argon's: $2 per million input tokens, $10 per million output tokens, and a 95% discount on cached input. After the introductory period, pricing becomes $4 per million input tokens and $20 per million output tokens. This pricing is attractive to developers—provided they can get API access.

The problem now is access. Google says it will continue gathering feedback, iterating on guardrails, and then open it to developers, enterprises, and consumers as soon as possible. There is no specific date. Some commenters complain: they announce a model update but not the model itself. Others call Carbon a shiny rock and say the next one should be called Diamond. These reactions show that there is still a gap between ranking first on a leaderboard and being actually available.

Who Should Pay Attention

Teams building security tools and managing vulnerabilities should keep an eye on the Fairwind Program. Carbon's numbers for vulnerability discovery, validation, and remediation are already capable of changing workflows. For application developers, don't put Carbon into near-term plans yet; wait for API and pricing details to land. Ordinary users won't be able to use it in the short term, but capabilities like these will first enter browsers, cloud platforms, and security products, so the impact won't stop at the launch event.

By separating the release timing from the access scope, Google shows it wants to find a balance between safety and capability. For developers, the most practical questions remain: when can we use it, and will using it actually cost this price?

Sources: Google's official blog and X posts.