Google has made Gemini 4 Carbon available to paying developers through the Gemini API, introducing another high-end model for complex agent and coding workloads. Access is available through Google AI Studio and Android Studio, while enterprise access is being offered through Gemini for business.
The initial API price is $2 per million input tokens and $10 per million output tokens through November 30, 2026. Google says those rates will rise to $4 and $20 respectively after the introductory period. That unusually short pricing window means developers evaluating Carbon in November need to calculate production costs against the December rate, not only the launch price.
Carbon targets difficult, long-running work
Google is positioning Carbon for long-horizon agent workflows, complex coding, scientific research and cybersecurity tasks. It is also rolling out to Google AI Ultra subscribers through Antigravity, the Gemini app and AI Mode in Search.
The release is separate from Gemini 4 Hydrogen, Google's lower-cost model for high-volume and repetitive agent work. Together, the products point toward a workload-based model strategy: Carbon for tasks where additional capability may justify a higher price, and Hydrogen or other less expensive models for routine execution.
That distinction matters because an agent session is not necessarily tied to one model from beginning to end. A system might use a cheaper model for classification, planning or repeated tool calls, then escalate only difficult steps to Carbon. For teams operating multi-model API gateways, Carbon therefore creates a routing decision rather than merely another model identifier.
The December price change requires more than a catalog update
The scheduled increase doubles both published token rates after November 30. Any gateway, reseller or internal platform exposing Carbon must attach effective dates to its rate card and ensure that estimates, customer dashboards and invoices switch at the same boundary.
Evaluation results gathered under the introductory price also need careful interpretation. A workflow that appears economical at $2 per million input tokens and $10 per million output tokens may no longer meet its budget at the standard rate. Teams should compare cost per completed task, including retries and agent steps, instead of comparing token prices in isolation.
Routing policy is the practical response. Carbon can be reserved for requests that meet explicit complexity, risk or quality thresholds, while routine traffic remains on lower-cost models. Spend controls should also operate at the model and API-key level so that an unexpected routing change does not silently move a large workload onto the more expensive tier.
Billing implementations should preserve the exact model revision, requested model and applied price version for each call. They should also account for cached input under the applicable provider terms rather than treating every input token identically. This is particularly important for agents that repeatedly reuse long instructions or shared context.
Availability still needs deployment-level verification
Google's announcement establishes availability through the Gemini API for paying developers, but it does not establish identical access across every Vertex AI configuration, region or Google Cloud deployment surface. Organizations with residency, procurement or network requirements should verify the specific endpoint they plan to use before treating Carbon as a production option.
The same caution applies to Google's performance and safety claims. Vendor benchmarks can help determine what to test, but they do not prove that Carbon will outperform another frontier model on a particular codebase, scientific workflow or security task. A useful evaluation should reproduce real prompts, tools, context lengths and failure-handling behavior from the intended deployment.
Cybersecurity use also raises an operational issue beyond model quality. A model assigned to inspect systems, write exploit demonstrations or operate tools should run with restricted credentials, explicit approvals and auditable actions. Higher capability does not remove the need for least-privilege controls around the agent.
Who needs to act now
Direct API users should confirm access, benchmark Carbon against their actual workloads and model costs using the post-introductory rates. Gateway operators need a Carbon catalog entry, effective-dated pricing, model-level reporting and routing policies that distinguish difficult work from ordinary traffic.
Resellers have an additional margin question. Customer prices that merely pass through November's rate may become uneconomic in December unless contracts or rate cards already support provider price changes. Usage reports should make the applicable rate period visible so customers can reconcile a November test with a later production invoice.
Carbon expands Google's frontier-model offering, but its immediate operational lesson is straightforward: capability selection, billing and routing must be managed together. The launch price may encourage testing; the durable decision depends on measured task performance at the scheduled standard rate.