Google Unveils Gemini 4 Carbon With Focus On Coding And Cybersecurity
Google says its new AI model leads several benchmarks, but access is initially limited to selected cybersecurity defense companies.
Google has introduced Gemini 4 Carbon, a frontier AI model it says improves on its recently released Gemini 4 Argon in coding, longer-running tasks and cybersecurity. Google’s reported benchmark results put Carbon ahead of leading models from OpenAI and Anthropic in several areas, though access is initially limited to selected cybersecurity defense companies.
Table of Contents
What Google is claiming
Carbon ranks first on the Artificial Analysis Intelligence Index v5.0, according to both Engadget and The Verge. Engadget reports that Google’s score for Carbon is 59, ahead of Claude Opus 5.5 and GPT-6 Astra. The outlet says Google also claims strong results across agentic coding, computer use, knowledge work, science and math.
The Verge reports that Carbon leads two cybersecurity benchmarks: SEC-bench Pro, which tests whether models can find real-world bugs, and ExploitGym, which assesses their ability to turn vulnerabilities into exploits. Google says Carbon is better at identifying and prioritizing vulnerabilities and addressing their underlying causes. These are Google’s claims about its model and benchmark performance, rather than results from broader deployment.
Security, reliability and longer tasks
Beyond benchmark scores, Google says Carbon is designed to stay with complex engineering work for longer. According to The Verge, the company says its internal teams have seen the model continue working for days through setbacks on multi-step tasks.
Google also claims improvements in how Carbon behaves while carrying out those tasks. Engadget reports that the company highlighted stronger resistance to prompt injection, fewer shortcuts and loopholes, better collaboration, and a greater ability to follow through without overstepping. The Verge says Google claims Carbon produces half as many unsupported claims as Argon. Engadget reports a separate comparison: Google says Carbon is twice as likely as Argon to generate secure code.
Access and implications for tech teams
For now, the model is available only to a selected group of cybersecurity defense companies. Both outlets say Google plans to extend access to other trusted partners through its Fairwind Program. Engadget adds that Google intends to release Carbon more widely later; neither report gives a date for that release. The Verge says Carbon will have the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens.
For software engineers and security teams, the immediate significance is therefore limited by access. The reported gains point to potential uses in coding, vulnerability discovery and remediation, but most teams cannot yet evaluate Carbon in their own workflows. The reports do not establish how its benchmark results will translate into day-to-day work or hiring in the US or Israel.
Sources
This story was compiled by AI from the reports below. Read the originals for the full details.