Google releases Gemini 4 Carbon to paying users and unveils two new TPU chips

Carbon reaches paying customers

Google is rolling out Gemini 4 Carbon to the general public on Tuesday, 10 November, after a limited release to a group of cybersecurity defenders in late October. Google calls Carbon its strongest model and says it can compete with the best systems from Anthropic and OpenAI. The company says Carbon pushes the frontier further than Argon, the Gemini 4 model it launched last month, but Argon remains the better choice for most software engineering and knowledge work. Carbon is built for harder tasks such as multi-step coding and open-ended scientific research. Paying developers and enterprises can use it through the Gemini API in Google AI Studio and Gemini for business, while AI Ultra subscribers, who pay $100 a month, also get it in the Gemini app, AI Mode in Search and Antigravity. The timing for lower-tier and free users is unclear.

Pricing and benchmarks

The introductory API price is $2 per million input tokens and $10 per million output tokens. Those prices double at the end of November, when the lower-price introductory period finishes. Per token, Carbon remains well below frontier rivals, although models can use different numbers of tokens for the same task, so total costs may not follow the per-token gap. Google's launch benchmarks place Carbon among the most powerful models on most metrics, behind Anthropic's Fable 5.2, which launched on 9 November. Those results come from Google's own announcement, and companies can choose benchmarks that show their tools in the best light.

Input tokens2

Output tokens10

Input tokens

2 $ per million tokens

Output tokens

10 $ per million tokens

Data from the article text

Swarm agents and science tools

Google is previewing a swarm feature on its Antigravity agentic development platform that turns a single prompt into a coordinated team of agents, aimed at long-horizon work. Google says early Carbon users helped uncover zero-day vulnerabilities in a short span of time. The company also reports advances in protein structure prediction, a capability needed for developing medications. Google is giving 10,000 scientists free access to Carbon for six months and has already onboarded thousands of them. Engadget notes that the program could also help train Google's models in turn.

Two new chips for cloud and confidential AI

On Monday, 9 November, Google unveiled TPU 8tx and TPU 8i Max, two variants added to the eighth generation of its tensor processing units introduced in April 2026. TPU 8tx, designed with AMD and based on the TPU 8t chip, is geared toward confidential training and running AI models, with security built into the hardware layer. It also includes an inferencing network that lets it run reasoning alongside training, so it can distill and train small models on the fly without moving proprietary data between devices. TPU 8i Max combines two TPU 8i chips from April into a single package, doubling memory capacity, processing power and bandwidth. The launch comes as Microsoft, Amazon, OpenAI and Anthropic also build their own AI chips to reduce reliance on Nvidia, while Google seeks to lessen its dependence on chips made by other companies for its cloud business.

Gemini Robotics 2.5 in the factory

Google DeepMind revealed Gemini Robotics 2.5 and a new open source safety benchmark called ASIMOV-jailbreak. The platform pairs Gemini Embodied Reasoning, which plans tasks from instructions and live video, with a vision-language-action model that generates robot movements. In evaluations across more than 200 activities, including serving food with tongs and using a pipette, the model achieved a 30 to 100 percent higher success rate than the previous version and runs 20 to 30 percent faster, though those results come from laboratory testing. In pilots with Apptronik and Mercedes-Benz, Apptronik's Apollo 2 humanoid kitted automotive parts in a few European factories with a 96 percent success rate. Kanishka Rao, a DeepMind director, said closing the remaining 4 percent gap will be a steep road.