Aravind Srinivas shares early Gemini 4 Carbon evaluations and reports WANDR SoTA

What happened

Aravind Srinivas shares early evaluations of Gemini 4 Carbon, calling it the strongest model they have seen for long-form research and answer quality, with a new SoTA on the Wide-And-Deep-Research (WANDR) benchmark. He says it is coming soon to all Perplexity Pro and Max users inside Computer, but the post cuts off while describing an availability condition.

Read next

• More changes in today’s briefingMistral releases improved Mistral Large 4 open weights for coding and cybersecurity

The story so far

Google DeepMind introduced Gemini 4 Carbon, which topped Artificial Analysis’s Intelligence Index v5.0 with a score of 59, while access remains limited to selected cybersecurity defenders. Google plans to expand access through its Fairwind Program before offering Carbon to developers, businesses, and consumers after gathering feedback and adjusting safety guardrails. Artificial Analysis said Carbon led four of nine evaluations, but scored 59 in its coding-agent comparison using Antigravity CLI, behind Claude Opus 5.5 at 62. Google reported cybersecurity scores of 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench, all evaluated without guardrails. Google also said Carbon was twice as likely as Argon to generate vulnerability-free code. In early Perplexity evaluations, Aravind Srinivas called Carbon the strongest model they had seen for long-form research and answer quality and reported a new leading result on the Wide-And-Deep-Research (WANDR) benchmark. Google set introductory pricing at $2 per million input tokens and $10 per million output tokens; Artificial Analysis described this as a 50% launch discount. The restricted rollout drew criticism from Nathan Lambert, who objected to announcing scores for a model he could not publicly use. He later softened his criticism, citing difficulties he had heard about in getting a frontier model released.