Gemini 4 Carbon is open to paying developers and enterprises, and its price doubles on 1 December

On 10 November 2026 Google made Gemini 4 Carbon available to developers, enterprises, scientists and consumers, and called it its strongest model. Two weeks earlier, Carbon had gone only to a set of cybersecurity defenders through Google's Fairwind Program.

Developers reach Carbon through the paid tier of the Gemini API, in Google AI Studio and Android Studio, and it is not on the Free tier. Enterprises reach it in Gemini for business, through the API or the model menu, and in Google Antigravity and CodeMender.

Google AI Ultra subscribers get Carbon in Google Antigravity, the Gemini app and AI Mode in Search. Scientists can apply for six months of free access through Antigravity, subject to eligibility, capacity and selection.

A token is the small unit of text that API use is billed in. Until 30 November 2026, Carbon costs $2 per million input tokens and $10 per million output tokens, which Google calls an introductory price. From 1 December 2026 it costs $4 and $20.

Google's pricing page gives Gemini 4 Argon the same prices as Carbon on every tier, so Argon's Gemini API price rises on 1 December too. On the Batch and Flex tiers, both go from $1 and $5 to $2 and $10. On the Priority tier, both go from $3.60 and $18 to $7.20 and $36.

The same page says data sent to Carbon or Argon on the Standard, Batch, Flex and Priority tiers is not used to improve Google's products.

When Google first announced Carbon on 28 October 2026, it gave the same two prices but no end date for the introductory period.

A cost estimate built on $2 and $10 holds only until 30 November. Budgets for December onwards need $4 and $20, for Carbon and for Argon.

Teams already on Argon through the Gemini API pay twice as much per token from 1 December without changing a line of code.

As an illustration of ours, take an agent that uses 10 million input tokens and 2 million output tokens a day. At the Standard rate it costs $40 a day until 30 November and $80 a day from 1 December.

Anthropic lists Claude Sonnet 5.5 at $2 and $10 and Claude Opus 5.5 at $4 and $20. By our comparison of list prices, Carbon matches Sonnet 5.5 today and Opus 5.5 from 1 December. The cost of a task still depends on how many tokens each model uses, which we have not measured.

Google's documentation says Carbon matches Argon on context and price. Both handle up to 1,048,576 tokens in total, shared between the prompt, the model's thinking and its answer. Both offer three thinking levels, low, medium and high, with medium as the default.

Google says Carbon spends more thinking tokens per turn than Argon, so that it can explore a problem and check its work. That raises token use and response time. Google recommends Argon for most software engineering and knowledge work, and Carbon for harder tasks such as complex multi-step coding, debugging and open-ended scientific research.

Thinking tokens are billed as output, at $10 per million now and $20 from 1 December. Google's docs say every token the model processes is billed. That includes the instructions it uses internally when it searches Google to ground an answer. On the Standard and Priority tiers each search is billed too, with 5,000 a month free across Gemini 3 and newer models, then $14 per 1,000.

Cached input, where a repeated part of a prompt is stored and reused, costs $0.10 per million tokens until 30 November and $0.20 from 1 December. Google applies it by default to prompts of at least 4,096 tokens.

Code moving from Argon needs only the new model name, plus room for longer thinking in its output limits and timeouts. Code moving from Gemini 3 or earlier needs more work. The older way of caching prompts by hand is not supported, and the settings that control how random an answer is are ignored. A standard reply stops at 65,536 tokens.

Google says hard prompts can take several minutes. Its docs advise timeouts of at least 15 minutes, or background execution, which runs the work on Google's side. For teams that don't want their interactions stored, the docs point to continuing the work on the client side.

Google says Carbon ranks first on three benchmarks run by other companies: CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench. By Google's account it does so at under half the cost per task of the closest model from another provider, a claim we have not tested.

Google says its safeguards blocked 99% of attack attempts in its own simulated cyberattack tests. It also says resistance to prompt injection, where hidden instructions sit in content the model reads, still needs work.

Because Carbon and Argon cost the same per token, the price list says little about which is cheaper. Carbon is likely to cost more per task wherever it thinks for longer, and Google gives no figure for how much longer.

Google's advice to test at all three thinking levels and pick the lowest that meets your quality bar also works as a cost control. Each step down can mean fewer output tokens, which cost $20 per million from December.

Google says hard prompts can take several minutes and that Carbon can spend more of the output limit on thinking. A short timeout or a low limit can cut a request off before the answer is complete.

Google's docs point teams that don't want interactions stored to the client-side option. Whoever signs off data handling can choose between the two.

GitHub made Gemini 4 Carbon generally available in GitHub Copilot on 10 November 2026, for Copilot Pro+, Max, Business and Enterprise, in a gradual rollout.

Carbon in Copilot is billed at the provider's list price under usage-based billing, and GitHub's notice gives no multiplier.

Business and Enterprise admins control access through the model policy in Copilot settings. Under default model enablement, new models switch on automatically, unless an admin has turned off the global default or disabled this model.

Where the default was left on, developers can start using Carbon as the rollout reaches them, with no admin step. Each request is then billed at list price.

GitHub bills Carbon at the provider's list price, so the 1 December rise may reach Copilot bills. GitHub's notice does not say.