# Gemini 4 Carbon Is on pirouter: Price, Endpoints, and Where It Fits

> Gemini 4 Carbon is available on pirouter at $4 per million input tokens and $20 per million output tokens. See its supported endpoints, thinking settings, and when to use it instead of Argon.

- Published: Nov 11, 2026
- Author: PiRouter Team, Engineering
- Tags: models, pricing, cost
- Canonical: https://pirouter.ai/blog/gemini-carbon-public-access-pricing

---
**Gemini 4 Carbon is now listed and available on pirouter as `google/gemini-4-carbon`.** Its [pirouter model page](/models/google/gemini-4-carbon) lists **$4 per million input tokens, $20 per million output tokens including thinking, and $0.20 per million cached input tokens read**. This is pirouter's published price—not a temporary price quoted in Google's direct API announcement. Use Carbon for difficult multi-step work worth testing against an acceptance standard; Google still positions Argon as the better starting point for most software and knowledge work.

## Where can you call Carbon on pirouter?

The [public endpoint listing](https://api.pirouter.ai/v1/models/google/gemini-4-carbon/endpoints) identifies three supported shapes: **Chat Completions**, **Responses**, and **Gemini `generateContent`**. The model's public ID is `google/gemini-4-carbon`, with `gemini-4-carbon` also listed as an alias in the endpoint data. The endpoint listing does **not** list Anthropic Messages for this model. If your application depends on Messages-specific request and response fields, don't infer compatibility merely from the presence of another model in the catalog; choose one of Carbon's listed endpoints and verify your client integration.

The [pirouter model details](/models/google/gemini-4-carbon) give a **1,048,576-token context window**—about 1M tokens—and a **65,536-token maximum completion**. Inputs include text, images, audio, video and files including PDFs; output is text. Reasoning is on, with **low, medium and high** effort settings and **medium** as the listed default. These are the model's published limits and settings, not proof that every long document will be understood equally well or that every caller has the same account quota.

That separation matters when you move an existing workflow. A model can accept a PDF and still need document-specific citation checks. A tool-using agent can have enough context on paper and still fail a deadline through long reasoning or repeated calls. Check your account's actual access and limits, select the endpoint your client implements, then test the path with a representative request before replacing a working default.

## What does the published price cover?

pirouter's [public model catalog](https://api.pirouter.ai/v1/models) lists the following rates for this model. All figures are USD per **one million billed tokens**, not per request:

| pirouter billing component | Published rate |
|---|---:|
| Uncached input | $4 |
| Output, including thinking | $20 |
| Cached input read | $0.20 |

The cache-read rate applies to billed cached input reads; it is **not** a promise that every request will hit cache. Raw API input totals may already include cached tokens. Calculate cost from the billing categories: uncached input and cached reads use their respective rates; do not add cached reads to an input total that already contains them. Output tokens used for thinking matter when estimating the cost of a difficult task. Google's announcement can help explain Carbon's positioning, but its direct-API promotional pricing is a **different commercial offer** and is not the price pirouter is publishing. Do not transfer another channel's discount or access entitlement to this one.

A useful budget records each call's billed input, cached-read input and output separately, then adds retries and tool-related work across the whole task. If you are comparing against Argon, take the actual bill and accepted result for the **same** task—not a per-million-token price divided by a benchmark score. We have not run a controlled Argon-versus-Carbon evaluation or measured a typical Carbon token count for your workload.

## When should Carbon replace Argon?

[Google's November availability announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/) says Argon remains suitable for most software engineering and knowledge work; Carbon is meant for unusually demanding coding, debugging and open-ended research. Google says Carbon can use more tokens to explore and verify. Its [model card](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Carbon-Model-Card.pdf) also notes possible slow responses, timeouts and hallucinations, particularly relevant to long-running workflows. These are Google's descriptions, not a claim that we independently verified a fixed quality or cost advantage.

For a trial, pick a task with a concrete failure mode—say, a cross-package migration whose tests must pass or a set of dense documents whose figures must trace to sources. Keep the input, tools, deadline and review criteria fixed. Log accepted outcomes, full-call token usage, retries, elapsed time and human corrections for Carbon and the current Argon path. If both pass, a heavier reasoning path may buy nothing; if Carbon removes costly reruns, a higher per-call bill may still be worthwhile. Decide from that result, then keep a rollback path.

Finally, the publicly listed pirouter model is the ordinary `google/gemini-4-carbon` path. Google's specialized **Carbon for cyber** access is handled separately through its [Fairwind Program, as described in the launch announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/). pirouter's public listing is **not** a claim to offer that specialized defender-only version, nor does an Antigravity `/swarm` preview become a built-in feature of every API call. Check the [model page](/models/google/gemini-4-carbon) and the endpoint listing for the specific service you can actually use.

---

## Sources

### blog.google

- [Google's November availability announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/)

### storage.googleapis.com

- [model card](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Carbon-Model-Card.pdf)