Gemini 4 Carbon debuts at #1 on Mercor's APEX-Agents leaderboard
From the AI digest for Oct 31, 2026
What happened: Gemini 4 Carbon debuted at #1 on APEX-Agents, a leaderboard testing models on multi-step tasks drawn from investment banking, management consulting and corporate law — and it is the only model on the board with a mean above 90%, Demis Hassabis's repost of Mercor's post says. Carbon scored 84.2% Pass@1 and a 90.4% mean, ahead of Gemini 4 Argon (82.4%, 87.4%) and Claude Sonnet 5.5 (75.5%, 83.7%). Carbon leads management consulting at 91.9% Pass@1 and investment banking at 86.9%, and takes #2 in corporate law at 73.8%, where Argon still leads at 76.7%.
What's genuinely new: A #1 debut on a new agentic leaderboard for Google's second Gemini 4-generation model — the board's first above-90% mean — with a vertical-by-vertical split showing where it beats and trails its own predecessor, arriving via Carbon's own team's amplification of the vendor's board, not an independent writeup.
Watch: Independent checks of the numbers, whether Carbon's lead holds as other labs run the board, and how it plays against Anthropic's response to being placed third on a vertical its own Opus line previously led.
Links: Demis Hassabis on X · Oct 30, 1:52 PM ET