Google's Gemini 4 Carbon ranked first on the MathArena leaderboard on November 13, displacing OpenAI models from the top position, while the company simultaneously released Gemini 4 Hydrogen, a faster model for everyday coding and agentic tasks.
The MathArena result gives Google an independent performance marker on competition-level mathematics at a moment when the company is broadening the Gemini 4 family beyond its frontier tiers.
Carbon was released on November 10, as Automatica reported at the time7, positioned for long-horizon coding, scientific reasoning, and cybersecurity work, at an introductory API price of $2 per million input tokens and $10 per million output tokens. Earlier that week, Automatica also covered6 predictions that Carbon would outperform competing models on benchmarks, noting that no disclosed benchmark substantiated the comparisons.
The MathArena leaderboard1 evaluates large language models on recent mathematics competitions, olympiads, and research problems, using questions released after training cutoffs to limit benchmark contamination. Gemini 4 Carbon (high) posted an expected performance of 85.4% (±3.0%), ahead of Anthropic's Claude-Fable-5.2 (max) at 82.8% (±5.3%) and OpenAI's GPT-6.1 Sol (max) at 80.6% (±2.7%). Each model runs four times per problem; MathArena combines observed scores with an item-response-theory model to estimate performance on unanswered questions.
The ranking was noted publicly when Demis Hassabis retweeted a post by MathArena contributor Jasper de Koninck3 describing the result as the first time in a long time a non-OpenAI model had reached first place.
In a separate announcement on November 132, Google introduced Gemini 4 Hydrogen as a speed-oriented complement to Carbon and Argon. According to the blog post, Hydrogen surpasses Gemini 4 Argon on the Artificial Analysis Intelligence Index at its default high thinking level, scoring 51 against Argon's 50, and reaching 54 at its highest xhigh setting. Google said Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash at its default setting and more than triples Terminal-Bench 4.1 scores, from 18.2% to 57%. The announcement also cites Hydrogen ranking fifth on Agent Arena, first on Vals Tax Agent Bench (52.3%), tied for first on Vals Legal Research Bench (55.8%), and posting 71.6% on Agents' Last Exam v1.1. These figures are Google's own claims; the announcement includes no independent test results.
As Automatica reported when Hydrogen first launched5, the model is positioned for speed in everyday coding and agentic tasks alongside the reasoning-heavier Carbon and Argon.