Google launches Gemini 4 Carbon first for cyber defenders
Original title: [AINews] Gemini 4 Carbon tops leaderboards, launches for cyber defenders
DeepMind claims 88.2% on SEC-bench Pro; Artificial Analysis ranks Carbon first on Intelligence Index v5 at 59.
Google DeepMind announced Gemini 4 Carbon two weeks after Argon, aimed at cybersecurity and long-horizon tasks, with first access for trusted defenders through the Fairwind Program. Google cites 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. The recap says the model has a 1 million-token context and accepts text, image, video and speech, with text output. Independent listings put it first on several indexes, while pricing quotes differ and coding results were measured in Antigravity CLI.
Unconfirmed: A stronger unreleased Gemini and the identity of hidden “Obsidian 3” are unconfirmed, and listed prices disagree.
- Google says Terminal-Bench 4.0 cheating successes fell to 0.0% and attempts to 0.4%.
- Artificial Analysis lists $4/$20 standard pricing and a 50% launch discount with no end date.
- Phil Schmid quotes $2/$10 per million input/output tokens.
- Carbon’s hallucination rate rose from Argon’s 4% to 11%, still called second lowest.
- Sundar Pichai expects faster training and launches from recursively self-improving systems.
- Vendors
- Organizations
- Google DeepMindArtificial AnalysisFairwind Program
- AI models
- Gemini 4 CarbonGemini Argon
From the original article
14 min read at latent.spaceGoogle’s new frontier model leads independent evals, with access initially limited to trusted defenders.
By swyx .ai · latent.space
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.