Gemini 4 Carbon: introductory price and AI task cost

On November 10, 2026, Google announced Gemini 4 Carbon availability for developers, businesses, scientists, and consumers, with access depending on the channel. For inference buyers, an introductory rate equal to Argon's calls for a cost-per-accepted-task comparison before changing the route.

Key facts

• Carbon and Argon: US$ 2 per million input tokens and US$ 10 per million output tokens at the introductory rate.
• The introductory period ends on 30/11/2026; the exact time was not disclosed.
• After it ends, public-access Carbon: US$ 4 per million input tokens and US$ 20 per million output tokens.

How much does Gemini 4 Carbon cost, and until when?

Google reports an introductory rate of US$ 2 per million input tokens and US$ 10 per million output tokens. After that period ends on November 30, 2026, it announces US$ 4 and US$ 20, respectively, for Carbon intended for public access. The exact time was not disclosed.

Argon shares Gemini 4 Carbon's introductory rate, but the announcement does not disclose its later rate, so the change announced for Carbon cannot be extended to Argon. A rate is not a bill. Output costs five times as much as input per token in each tier, and Carbon's later rate doubles the introductory amounts without determining the buyer's expenditure, which depends on consumption, attempts, acceptance, currency conversion, and charges. Google says Carbon can use more tokens than Argon on complex tasks; parity per unit therefore does not settle the comparison of AI cost per task and token pricing. In the nominal component, I and O represent millions of input and output tokens, excluding cache, batch, and unverified specific conditions; the formula does not estimate local billing.

The nominal component is 2 × I + 10 × O dollars during the introductory period. For public-access Carbon after it ends, it becomes 4 × I + 20 × O dollars, keeping volumes equal solely to compare rates.

Where was Gemini 4 Carbon announced?

Google announces the Google Gemini API in Google AI Studio and Gemini for business for developers and paying businesses. For Google AI Ultra subscribers, it reports rollout in Antigravity, the Gemini app, and Search's AI Mode. The channels serve different audiences, and the announced rollout does not confirm individual access.

The announced availability of Gemini 4 Carbon follows the October 28 release to cyber defenders through Fairwind, documented in the earlier Carbon article, published on October 30, when there was no date yet for general access. These are distinct moments. In Antigravity, Google announced a preview of /swarm to coordinate agents, without confirming integration with Nexforce Router or availability in other channels, accounts, and regions. Google's model card describes Carbon as based on Argon, with a combined input and output window of up to 1 million tokens, not a limit for each direction. It accepts text, images, audio, and video and generates text. Specialized cyber capabilities require defenders verified through Fairwind. Dense or MoE architecture, parameters, quantization, and workload latency were not verified; confirm compatibility in the contracted channel before testing.

What can the benchmarks establish?

Google says Carbon leads CursorBench 4.0, WANDR, and AutomationBench with a cost per task more than two times lower than the next best external model in the cited comparisons. This is a Google claim about third-party tests. The full original methodology and independent reproduction of the comparison were not verified.

On November 12, 2026, CursorBench 4.0 displayed Gemini 4 Carbon High in first place, at 60.4% and US$ 9.22 per task, while AutomationBench showed the same configuration in first place, at 53.88% and US$ 2.12 per task. These are different tests. Cursor calculates average cost using input, cache read and write, and output prices, with a warning about variance; Zapier describes simulated environments, end-state criteria, and, in its Carbon report, a list-price estimate rather than a bill. Neither leaderboard proves its operators reproduced the comparison announced by Google, and WANDR was not consulted directly, preventing these readings from being treated as a complete reproduction of all three tests. Buyers should use the leaderboard to select candidates, not assume acceptance, latency, or expenditure for their workload without common tasks and criteria.

How is cost per accepted task calculated?

Cost per accepted task is the total known cost of a cohort divided by the number of accepted tasks in the same cohort. It includes unsuccessful attempts and known costs as well as the call producing an approved response. With no accepted tasks, the division is not calculated; with missing costs, the result is partial.

To measure cost per accepted task, the team defines the cohort and rubric before the calls, so each eligible model receives comparable tasks and criteria, including when human review is involved. Calls are not accepted tasks. Changing the rubric during the test changes the denominator of the routing decision; the numerator, meanwhile, combines known costs of cohort attempts, including failures, tools, and review when available. The formula is editorial analysis, not a result from Google or the benchmark operators, and the measurement is not the operation's complete bill because it is limited to known cohort costs. The figure relates tokens, attempts, and results, but cost includes rejected calls even when they do not lead to an accepted result: a rejected response remains in the cost and does not enter the denominator as an accepted task.

Observed cost per accepted task = total known cohort cost ÷ number of accepted tasks in the same cohort.

1. Choose representative tasks and define acceptance criteria before the calls.
2. Confirm model, channel, region, contract, and rate.
3. Record tokens, attempts, failures, review, latency, and costs.
4. Calculate the ratio using accepted tasks and identify missing expenses.

The analysis of routing when token prices change connects routing and measurement.

Frequently asked questions about Gemini 4 Carbon

The announcement answers what the public rate is, which channels were presented, and when the introductory phase ends. It does not prove access for every account, a universal benchmark result, or Carbon support in Nexforce Router. The answers below refer to the disclosed conditions and the limits of this reporting.

Google presents Gemini 4 Carbon to developers and paying businesses through the API and announces rollout to Ultra subscribers in other products, so account eligibility must be checked before scheduling calls. An announcement does not guarantee individual access. The cost comparison attributed to Google has a defined scope: the leaderboards consulted do not demonstrate reproduction of the original comparison or establish performance for every application, so they do not replace a buyer's own cohort with accepted quality and recorded attempts. The introductory period ends on November 30, with no exact time disclosed, and the announcement does not set Argon's later rate, preventing the announced conditions from being automatically transferred to another model. Nexforce Router catalog, price, support, and compatibility were not verified; the organization must confirm those conditions for its account because Google's announcement does not prove an offering on Router.

How much does Gemini 4 Carbon cost?

Google reports US$ 2 per million input tokens and US$ 10 per million output tokens at the introductory rate. After it ends, Google announces US$ 4 and US$ 20, respectively, for public-access Carbon. These are unit rates, not the contracted bill or cost per accepted task.

When does the introductory price end?

The period ends on November 30, 2026, according to Google; the exact time was not disclosed. After it ends, the announced rates of US$ 4 for input and US$ 20 for output per million tokens apply to public-access Carbon.

Is cost per task always lower?

No. Google's claim compares models in the cited tests, not across every workload. Benchmarks help select candidates, but a team must measure its own cohort, including the costs of rejected attempts and tasks accepted according to common criteria.

Is Carbon available on Nexforce Router?

Availability was not verified in this piece. Before routing calls, consult the catalog to confirm the model and the contract to understand the price and conditions applicable to the account. Also check support and compatibility in the channel: Google's public rate does not establish those conditions on Router.

Is the Google Gemini API available to every account?

The announcement presents the API to developers and paying businesses, but does not prove immediate or unrestricted access for every account and region. Before scheduling the test, the organization should confirm eligibility in the chosen channel and the conditions of its account.

References and further reading

Google's announcement is the source for the announced rates and conditions, while the model card describes the model and Cursor and Zapier publish their own leaderboards using different methods. The sources have limits. They do not demonstrate the complete cost comparison attributed to Google.

Google's November 10 announcement records availability and rates, consulted on November 12, 2026, while The Deep View reports US$ 4 and US$ 20 per million tokens and a 50% introductory discount, as editorial corroboration rather than a billing audit. The source matters. The model card documents the relationship between Carbon and Argon and the combined context window; CursorBench presents the result and average cost of its own test, with a variance warning, without allowing criteria to be transferred to another leaderboard without examining the method. Zapier publishes AutomationBench and a Carbon report that identifies the per-task estimate as list-price based rather than a bill, a distinction needed to avoid confusing evaluation with actual expenditure. Leaderboards can change; WANDR and the full methodology for the cost comparison cited by Google were not directly verified in this reporting.

• Google availability and pricing announcement.
• CursorBench 4.0, AutomationBench, and Zapier's Carbon report.

Next step: measure before changing the route

In Nexforce Router, the rate change is a reason to review the routing policy, not to replace a route immediately. The team should first verify eligible models and compare their tasks using common criteria. Carbon and Argon availability, price, and support on Router were not verified in this analysis.

Nexforce Router offers API routing with cost, performance, latency, and context criteria, call tracking, and spending limits, capabilities discussed in the article on how to prove real AI savings in production. Measurement comes before a switch. If Carbon and Argon are available and eligible on Nexforce Router, the team should calculate cost per accepted task using known costs and approved tasks from the same cohort, measuring before changing the route. Record the review date and channel rate, preserve the rubric and latency, and, after the introductory phase, recalculate input and output without assuming equal volumes or other charges. The previous policy should remain until there is enough evidence to change it, because the rate does not determine cost per task and no buyer-specific savings were demonstrated in this news report.