TL;DR

Google made Gemini 4 Carbon, which it calls its strongest model, generally available on 10 November. Paying developers and enterprises can use it through the Gemini API at the same introductory price as Gemini 4 Argon, and Google AI Ultra subscribers get it in the Gemini app, AI Mode in Search and Google Antigravity. The release follows two weeks in which only selected cyber defenders had access.

Where Carbon fits

Google positions Argon as the default for most software and knowledge work, and Carbon for the hardest tasks, such as debugging across many steps and exploratory scientific research. Carbon spends more tokens than Argon to reason and check its own work. It costs $2 for each million input tokens and $10 for each million output tokens.

Google says Carbon tops three workflow benchmarks (Zapier’s AutomationBench, CursorBench 4.0 and Perplexity’s WANDR) at under half the cost per task of the next best outside model. Alongside the launch it is previewing /swarm in Antigravity, which splits one prompt across a team of coordinated agents.

Cyber capability and safeguards

Carbon’s headline strength is security. Google says that, inside its Fairwind defender programme, Carbon leads benchmarks for finding, exploiting and patching vulnerabilities. Wiz used it to find previously unknown flaws at two major SaaS providers in under two hours, according to Google.

That capability is why the safety section matters. Google reports that in simulated attack tests, Carbon’s safeguards blocked 99% of trajectories from an adaptive red-teaming agent, cutting the per-attempt breach rate to 0.8% from 3.1% for Argon. Carbon scored 92% on a 375-task agentic honesty test, where Argon managed 70%. Google also concedes that resistance to prompt injection “remains an area for continued improvement”.

What it means

We covered Carbon’s limited release to cyber defenders in October, and Argon’s general release at the same price before that. Carbon was also used in the DeepMind work on verifying the Signal protocol we reported on 10 November.

In our view, UK businesses should treat the benchmark claims as Google’s own until independent testing catches up, but the pricing is the practical news: frontier-tier capability at Argon’s rate lowers the cost of trialling it on hard coding and analysis work. Security teams will want to read the separate safety and security report before enabling it widely, given the model’s offensive potential.