Google Drops Gemini 4 Carbon: Smarter Than Your Ex and Way Cheaper

Google DeepMind rolled out its heavyweight model Gemini 4 Carbon following its initial restricted phase, packing a massive context window of 1,048,576 tokens and a default medium reasoning tier designed specifically for grueling agentic workflows. Sitting atop the fourth-generation lineup next to its sibling Gemini 4 Argon, this architecture leans heavily into deep self-correction loops, burning roughly 205 million thinking tokens on massive index sweeps compared to Argon's 113 million.

In benchmark showdowns, Carbon secures the runner-up spot behind Claude Fable 5.2 on Artificial Analysis v5.0.1 with 60 points, yet dramatically undercuts the leader by executing tasks at a fraction of the cost—roughly $7.76 against $26.54. While pure software engineering metrics like FrontierSWE 2 and OSWorld-2.1 show mixed wins against GPT-6 Astra and Claude Opus 5.5, the model dominates specialized domains like GraphWalks with a 96.2% success rate in maintaining logical threads across endless agent trajectories.

Cybersecurity serves as the crown jewel of this drop, with Carbon scoring 88.2% on SEC-bench Pro and independently tracking down critical zero-day vulnerabilities through automated frameworks like Wiz and Fortify. Alongside the core model release, Google introduced the experimental /swarm mode in Antigravity, deploying isolated git worktrees where a coordinating instance spawns parallel worker agents to handle massive multi-package migrations without git conflicts.

The financial math of intelligence has officially shifted from luxury prestige to ruthless bargain-bin efficiency, leaving traditional software engineering gasping for air.