Google broadened access to Gemini 4 Carbon on November 10, bringing its latest high-end AI model to paying developers, businesses and Google AI Ultra subscribers. The release moves Carbon beyond its initial distribution to selected cybersecurity defenders, while keeping specialized security access separate.
Google’s launch announcement lists introductory API prices of $2 per million input tokens and $10 per million output tokens. Those rates run through November 30. The stated prices afterward are $4 and $20 respectively, so teams evaluating a new workflow need to account for the change rather than budget solely around the launch offer.
Different routes into the model
Google’s separate developer announcement says Carbon is available through the paid API tier in Google AI Studio and Android Studio. Ultra subscribers can try it in AI Studio and Google Antigravity. The developer release emphasizes sustained tasks, including multistep software engineering and scientific research, rather than positioning the model only as a faster chatbot.
For consumer use, Google’s main announcement says Carbon is rolling out in the Gemini app and AI Mode in Search for Ultra subscribers. It also previews a coordinated-agent feature in Antigravity. This is a paid, tier-dependent rollout, not a statement that every Gemini user receives identical access.
Capability comes with operational limits
Google DeepMind’s model card describes Carbon as based on Gemini 4 Argon, accepting text, images, audio and video while producing text. Its combined input and output context can reach one million tokens. A context window is the amount of material a model can consider in an interaction, not a guarantee that it will use every detail correctly.
The same document lists hallucinations, occasional slow responses and timeouts among known limitations. It gives March 2026 as the knowledge cutoff, with some areas potentially less current, and notes that higher thinking levels can consume additional tokens. Specialized cybersecurity capabilities remain available to verified defenders through the Fairwind Program.
The model card also publishes benchmark comparisons, but those are not uniform wins across every task. Buyers should treat individual scores as evidence about the evaluated setup rather than proof that Carbon is the best choice for all work. Google itself positions Argon as suitable for most software engineering and knowledge tasks.
Price per token is not the entire bill
The Verge’s coverage highlights an important distinction in the new pricing: a cheaper token does not automatically mean a cheaper completed task. Different models may generate different amounts of intermediate work before producing a useful answer.
For an organization choosing between models, that makes a small test on representative jobs more informative than the headline rate alone. The relevant comparison is the cost and time needed to obtain an acceptable result, including review and corrections. The November 30 pricing deadline provides a concrete point to revisit any early cost estimates.
