Gemini 4 Carbon

972 points · 775 comments on HN · read original →

Points and comments are a snapshot, not live.

Google announces Gemini 4 Carbon, a frontier model with state-of-the-art cybersecurity capabilities.

Google introduced Gemini 4 Carbon as its strongest model, ranking #1 on the Artificial Analysis Intelligence Index (v5.0) with a score of 59. It sets new benchmarks in cybersecurity across vulnerability detection (SEC-bench Pro: 88.2%), exploitation (ExploitGym: 54.8%), and patching (CWE-bench: 71.0%). The model shows large improvements in coding and long-horizon agentic tasks, scoring 1768 on WebDev Arena. Internal teams report a 2x uplift in secure code generation over Argon and significant reductions in model misalignment and fabrication. Carbon is being released initially to a subset of cyber defenders under the Fairwind Program, with broader access to follow. Pricing is $2/M input tokens and $10/M output tokens during the introductory period.

What commenters are saying

Commenters expressed cautious interest, calling the benchmarks promising but noting the model is not yet widely accessible. Some praised the progress on reducing shortcut attempts and fabrication-calling alignment improvements the most important aspect for usability. Others criticized Google's practice of announcing models with delayed public release, calling it vaporware or uncompetitive. Skepticism was voiced about the credibility of AI benchmarks, especially with private tests included in the AA index. A few commenters highlighted the aggressive introductory pricing as a competitive move. Sentiment is split between excitement over capabilities and frustration with limited access.