Google’s Gemini 4 Carbon debuted at No. 1 on Mercor’s APEX-Agents leaderboard with an 84.2% Pass@1 score and a 90.4% mean, becoming the only evaluated model to average above 90% across multi-step tasks in investment banking, management consulting and corporate law. Carbon beat second-place Gemini 4 Argon at 82.4% Pass@1 and 87.4% mean and third-place Claude Sonnet 5.5 at 75.5% Pass@1 and 83.7% mean, taking first place in management consulting at 91.9% and investment banking at 86.9% while placing second in corporate law at 73.8%, 2.9 percentage points behind Argon’s 76.7%.

Perplexity CEO Aravind Srinivas also said early evaluations showed Gemini 4 Carbon set a state-of-the-art score on the Wide-And-Deep-Research (WANDR) benchmark and called it the strongest model tested so far for long-form research and answer quality. He said Perplexity will roll out Carbon to all Pro and Max subscribers inside Computer once it has sufficient token capacity. Meanwhile, TestingCatalog reported that Google has been testing a Carbon-H build on Arena under the alias Argon 4 since Oct 31, while Google engineer Sumit Bagri noted Carbon scored 64.8% on Terminal-Bench-Science and cut cheating attempts on Terminal-Bench 4.0 to 0.4%.