- Gemini 4 Carbon scores 59 on the Artificial Analysis Intelligence Index v5, up from Argon’s 50, taking the top overall spot.
- Google built Carbon for multi-day, autonomous coding workflows and the full cyber defense lifecycle.
- Carbon still trails Claude Opus 5.5 on Artificial Analysis’s Coding Agent Index, 59 to 62.
- Because of its cybersecurity capabilities, Carbon is going through voluntary US government pre-release testing before wider access.
What Happened
Google DeepMind released Gemini 4 Carbon on Wednesday, just two weeks after shipping Gemini 4 Argon. Carbon tops the newly updated Artificial Analysis Intelligence Index v5 with 59 points versus Argon’s 50, and leads in four of the index’s nine individual evaluations, with particular strength in long-context and security tasks.
Why It Matters
A two-week gap between flagship releases signals how fast Google is now iterating against rivals like Anthropic and OpenAI. Google says Carbon ran unsupervised for multiple days on internal tests, working through dead ends on its own while handling multi-step tasks like root-cause analysis in large codebases.
Technical Details
On Gray Swan’s indirect prompt injection benchmark, Carbon cut the attack success rate from 0.7% with Argon to 0.0%. On SEC-bench Pro, which measures autonomous vulnerability discovery, Carbon scored 88.2%, leading all tested models; on ExploitGym, which tests turning vulnerabilities into working exploits, it reached 54.8%. In agentic coding on Artificial Analysis’s Coding Agent Index, Carbon still trails Claude Opus 5.5, 59 to 62. Pricing holds steady at Argon’s rate: $2 per million input tokens and $10 per million output tokens.
Who’s Affected
Google is rolling Carbon out gradually through its Fairwind Program, giving select cybersecurity experts first access as part of the US government’s voluntary pre-testing process, given the model’s vulnerability-discovery capabilities. Broader developer, business, and consumer access will follow once Google adjusts safety guardrails based on that feedback.
What’s Next
Independent evaluators including Vals AI and Arena.ai have corroborated Google’s benchmark claims, placing Carbon competitively against Claude Opus 5.5 and GPT-6 Astra at lower per-task cost. Whether Carbon’s security-testing access model slows real-world adoption compared to a full consumer launch remains the open question.
Source: The Decoder, Oct 28, 2026
Related: Gemini 4 Argon’s launch two weeks earlier and DeepMind’s cyberattack test results for Argon.
