How Much To Run AI
Restricted preview

Gemini 4 Carbon: Access, API Prices and Long Task Costs

Gemini 4 Carbon is Google DeepMind’s frontier model for sustained software engineering and cybersecurity defense. Google is opening it first to selected cyber defenders, with trusted partners following through Fairwind. Announced introductory prices are $2 per million input tokens and $10 per million output tokens; after that period, the rates rise to $4 and $20. Use Argon’s available API if you need a public Gemini endpoint today.

Content reviewed:

Gemini 4 Carbon announced token rates: introductory input $2 and output $10 per million; after the introductory period $4 and $20. Access starts with selected cyber defenders.
Announced token rates; access begins with selected cyber defenders. The introductory end date is not specified in the Carbon announcement.

Use this model

Online access and trials

Where does Gemini 4 Carbon fit?

  • Investigate and patch difficult bugs across a large repository with an agent that can run tests.
  • Evaluate vulnerability detection and remediation on systems your organization is authorized to assess.
  • Compare the complete token bill and accepted changes for long engineering tasks.

Who can use Gemini 4 Carbon?

Google announced Carbon on October 28, 2026. The first access goes to a subset of cyber defenders, followed by additional trusted partners through the Fairwind program. The announcement describes broader developer, business and consumer availability as a later step.

Start with the Fairwind eligibility and contact route if your organization fits that program. For an approved integration, use the endpoint, model identifier and quotas supplied with your access. The announcement does not supply a public Carbon API ID or a general signup endpoint.

For a project that must ship now, the Gemini 4 Argon guide covers its public API route. Carbon requires its own access approval, separate from an Argon subscription or an Arena demonstration.

Gemini 4 Carbon API pricing

Google’s October 28 announcement quotes USD token rates for Carbon’s launch. Cached input receives a 95% discount from the input rate. These are announced model prices; the initial access agreement determines what an approved user can actually call and bill.

The footnote doubles the standard input and output rates after the introductory period. The Carbon announcement does not give that period’s end date. Keep its introductory and later prices separate rather than importing Argon’s calendar.

Billed tokensIntroductory USD / 1M tokensAfter introductory period
Uncached input$2$4
Cached input$0.10$0.20, applying the announced 95% discount
Output$10$20

Estimate a complete agent task

For an assumed task using 1 million uncached input tokens and 100,000 billed output tokens, the introductory model charge is $3: $2 input plus $1 output. At the later rates, the same usage costs $6. This example assumes no separate tool, storage or infrastructure charges.

If 800,000 of those input tokens qualify for the cached rate and 200,000 remain uncached, input costs $0.48 and the total becomes $1.48 at introductory rates. Check your access terms for cache eligibility, retention and any additional service charges.

A long agent run makes many calls. Add usage across the entire run, including retries, repeated repository context and unsuccessful attempts. Compare spending per accepted patch or completed migration alongside elapsed time. Equal token prices do not establish equal task cost.

Artificial Analysis reports $7.76 per Intelligence Index task for Carbon (high), calculated at the standard $4/$20 rates rather than the launch discount. It reports $7.68 for Argon at the same standard rates. This is an evaluation-specific weighted cost, not a quote for an individual coding job or a public customer bill.

What do the Carbon benchmarks measure?

Artificial Analysis independently evaluated Carbon with high reasoning on October 28. It scores 59 and ranks first on Intelligence Index v5, while Carbon with Antigravity CLI scores 59 on the separate Coding Agent Index, behind Opus 5.5 with Claude Code at 62. The overall index lead does not make it the leading coding-agent setup.

Artificial Analysis lists a 1M-token context and text, image, video and speech inputs with text output. Its model page also marks Carbon as having no publicly available API; the benchmark listing does not provide a public access route.

For cybersecurity, Google reports 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. Its announcement says these external benchmark evaluations run without guardrails. Production access has safeguards, so the benchmark setup and an approved service deployment are different conditions.

Google also describes internal agents persisting for days on engineering problems, including difficult bug fixes. That illustrates the intended workload, but the announcement gives no corresponding token totals or cost per successful task. Use your own repository and acceptance tests when deciding whether a longer run pays for itself.

Choose Carbon, Argon or a local model

Carbon is relevant for organizations with approved access and engineering or defense work that benefits from sustained tool use. Evaluate whether it produces an accepted result with fewer interventions, and retain a record of the whole run’s cost.

Argon offers an available Gemini route for teams that need to start integrating now. A small local model such as LFM3 suits bounded on-device tasks when local processing matters more than frontier engineering capability. Those choices have different operating costs: hosted token usage for Carbon or Argon, and device memory, energy and runtime for local inference.

For interactive generated environments rather than repository work, Genie 4 is a separate world model with its own preview-access path.

Before you use it

  • Carbon access is restricted to the rollout’s approved organizations; public product availability is a separate release step.
  • The announcement provides no downloadable weights or hardware specification for self-hosting.

How to get started

Check Fairwind access for an eligible organization. For an immediate public Gemini integration, follow the Argon guide and keep its model ID and billing separate from Carbon.

Sources

Sources (5)

Calculators