Gemini 4 Carbon API Guide: Model ID, OpenAI SDK, Billing

As of the 2026-11-11 catalog, Router One lists Gemini 4 Carbon as google/gemini-4-carbon. The gateway also accepts Google's own id, gemini-4-carbon, for the same catalog entry. Send "model": "google/gemini-4-carbon" to https://api.router.one/v1/chat/completions with a Router One key — or point any OpenAI-compatible SDK or client at https://api.router.one/v1 — and the request is routed, metered and logged like every other model. As of the 2026-11-11 plan response, no tier of the Pro, Max or Ultra plans lists Gemini 4 Carbon, so every call bills per token to the wallet at the rates on the Gemini 4 Carbon model page, plan or no plan.

This guide covers what the catalog lists for the id, what Google documents about the model, how billing works, how it differs from Gemini 4 Argon, the first request with curl and the OpenAI SDKs, which clients can call it, and access from mainland China; every Google statement below is attributed and checked 2026-11-11, and the catalog and plan observations are dated. The live catalog is the source of truth for a new request.

Gemini 4 Carbon at a glance

What Google documents for Gemini 4 Carbon

Per Google's Gemini API release notes (November 10, 2026 entry; checked 2026-11-11), Gemini 4 Carbon (Google's model id gemini-4-carbon) is generally available on the Gemini API. Google's Gemini models page (checked 2026-11-11) describes it as "Our strongest reasoning model, built for complex software engineering, multimodal knowledge work, scientific reasoning, and cybersecurity defense."

Google's Gemini 4 Carbon model page (checked 2026-11-11) lists the following for Google's own Gemini API:

• Status. Google lists the model as a stable version, and Google's deprecations page (checked 2026-11-11) gives November 10, 2026 as its release date and no announced shutdown date.
• Token limits. A context window of 1,048,576 tokens that input and output share, the unified 1M-token context window that Google's What's new in Gemini 4 Carbon page describes (checked 2026-11-11), with a maximum output of 65,536 tokens by default. Google's long decode feature, which runs on Google's own APIs, lets a response go past 65,536 tokens up to the rest of that window.
• Inputs and output. Per Google, text, image, video, audio and PDF in, text out. The Router One catalog entry lists text and image input, so the samples below send only those.
• Features. Google lists function calling, structured outputs, code execution, search grounding, URL context and implicit context caching as supported (explicit caching is not), and image generation, audio generation and the Live API as not supported.
• Thinking. Google lists the levels low, medium and high. The next section has Google's default and the details.

Those are Google's statements about its own Gemini API; Router One is responsible for the catalog entry and for how the gateway routes and bills it. The catalog lists the tool-calling flag for this id, so function tools go in the standard Chat Completions tools array. Google's built-in tools are Gemini API features: on Router One's Chat Completions endpoint the hosted search tool is handled for google/gemini-3-flash only, and for every other id the tools array is forwarded as sent (API compatibility facts). Test the tool definitions your code sends on your own prompts before you move a workload.

Thinking levels and reasoning effort

Per Google's What's new in Gemini 4 Carbon page and its thinking guide (both checked 2026-11-11):

• Levels. Gemini 4 Carbon thinks at low, medium or high, and medium is the default; Google recommends medium for most tasks and suggests running your evaluations at all three levels, then picking the lowest one that meets your quality bar.
• Token use. Per Google, on hard problems Gemini 4 Carbon spends more thinking tokens per turn than Gemini 4 Argon to explore and verify its work, which increases per-turn token use and latency; to use fewer tokens, Google suggests lowering the thinking level.
• Output cap. Google's output cap counts thinking tokens as well as the answer, so a response that reaches it while the model is still thinking comes back truncated or empty; Google recommends raising an explicit output limit to at least 32,768 tokens.
• No off switch. Google's OpenAI compatibility guide (checked 2026-11-11) maps a reasoning_effort of minimal or low to the low thinking level for Gemini 4 models, medium to medium and high to high, uses the model's default level when the field is absent, and says reasoning cannot be turned off for Gemini 4 models.

Those rules describe Google's own API. The samples below do not set reasoning_effort; if your code sets one, run it on your own prompts before you move a workload. Thinking output is billed at the model's posted output rate, with no separate reasoning line (pricing facts), so thinking tokens add to a call's wallet charge.

How billing works

Gemini 4 Carbon bills per token at the input and output rates posted on its model page. Router One posts its own rates for each model, with pricing up to 90% off official prices on select models, and this guide prints no per-token figures because they go stale. As of the 2026-11-11 plan response, no tier of the Pro, Max or Ultra plans lists Gemini 4 Carbon, so every call bills the wallet at the posted rates, plan or no plan.

Google prices its own Gemini API separately: per Google's pricing page and its What's new in Gemini 4 Carbon page (both checked 2026-11-11), Gemini 4 Carbon is at an introductory price on the Gemini API through November 30, 2026, Google's standard pricing takes effect on December 1, 2026, Google's output price includes thinking tokens, and the model is not available on Google's free tier. Those dates belong to Google's price list; the rate Router One charges for the id is the one on its model page.

Dashboard → Logs shows each call's model, tokens, wallet charge or plan quota used, latency and status (per-request observability), so you can add up a task's wallet charges from the calls it made rather than estimate them from rates.

Do subscription plans cover Gemini 4 Carbon?

No, as of the 2026-11-11 plan response: no tier of Router One's Pro, Max or Ultra plans lists Gemini 4 Carbon, so on a plan its calls bill the wallet per token at the posted rates, as they do without one. Gemini 4 Argon is a separate catalog entry: as of the 2026-11-11 plan response, it is in the Premium models tier of the Pro, Max and Ultra plans. Plan model lists change; the pricing page lists each plan's allowance and the live model lists.

Gemini 4 Carbon vs Gemini 4 Argon

Google's Gemini models page (checked 2026-11-11) describes Gemini 4 Argon as "Our efficient Gemini 4 frontier reasoning model, delivering faster responses and lower token spend per turn compared to Gemini 4 Carbon across complex coding, agentic, math, and science tasks." Per Google's What's new in Gemini 4 Carbon page (checked 2026-11-11), the two models share the same context window, built-in tools and per-token price on Google's own Gemini API, and the main difference is how many thinking tokens each spends per turn; Google suggests Gemini 4 Carbon for complex root-cause debugging, architectural design, dense multimodal synthesis, scientific proofs and security auditing, keeping Gemini 4 Argon where faster responses and lower token spend per turn matter, and says that moving from Argon to Carbon needs no request schema changes on its API.

On Router One, both are listed as of the 2026-11-11 catalog and both are served on /v1/chat/completions only, so the key, base URL and endpoint stay the same: on the wire, the switch is the model string, from google/gemini-4-argon to google/gemini-4-carbon. The plan side differs. As of the 2026-11-11 plan response, Gemini 4 Argon is in the Premium models tier of the Pro, Max and Ultra plans, so on a plan its calls draw that tier's allowance. As of the 2026-11-11 plan response, no tier lists Gemini 4 Carbon, so its calls bill the wallet per token, plan or no plan. Compare the posted rates on the Gemini 4 Argon and Gemini 4 Carbon model pages and run your own prompts on both before you move a workload.

The Gemini 4 Carbon vs Gemini 4 Argon comparison page renders their spec sheets and live rates side by side, and the Gemini 4 Argon API guide covers Argon's model id, endpoint and plan coverage.

Send the first request

1. Create a key. Dashboard → API Keys → Create Key. Keys look like sk-.... For a trial, give the key a maxSpend cap. A key's spend cap (maxSpend) limits what that key can charge to your wallet: once its cumulative wallet charges reach the cap, its wallet-billed requests return HTTP 402. Requests covered by your plan's quota draw from the account's shared plan quota and do not count toward any key's cap. Your other keys keep working (per-key cost tracking).
2. Keep a wallet balance. Gemini 4 Carbon calls bill the wallet, plan or no plan: no tier of the 2026-11-11 plan response lists it — top up in Dashboard → Deposit.
3. Set the base URL to https://api.router.one/v1 in any OpenAI-compatible SDK or client.
4. Call the model. The examples use the catalog id google/gemini-4-carbon and set max_tokens explicitly.

Chat Completions with curl, streaming:

curl https://api.router.one/v1/chat/completions \
  -H "Authorization: Bearer sk-your-api-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemini-4-carbon",
    "max_tokens": 32768,
    "stream": true,
    "messages": [
      {"role": "system", "content": "You debug concurrency problems. Give the root cause first, then the fix."},
      {"role": "user", "content": "Two workers retry the same payment after a timeout, and some customers are charged twice. Where is the race, and how do we make the retries safe?"}
    ]
  }'

The same request without streaming, from the OpenAI Python SDK (pip install -U openai) — only base_url and the key change:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.router.one/v1",
    api_key="sk-your-api-key",
)

response = client.chat.completions.create(
    model="google/gemini-4-carbon",
    max_tokens=32768,
    messages=[
        {"role": "system", "content": "You debug concurrency problems. Give the root cause first, then the fix."},
        {"role": "user", "content": "Two workers retry the same payment after a timeout, and some customers are charged twice. Where is the race, and how do we make the retries safe?"},
    ],
)
print(response.choices[0].message.content)
print(response.choices[0].finish_reason, response.usage)

The reply has the Chat Completions shape: the text is in choices[0].message.content, finish_reason says why generation stopped, and usage carries the input and output token counts.

Streaming from Node with the OpenAI SDK (npm install openai):

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.router.one/v1",
  apiKey: "sk-your-api-key",
});

const stream = await client.chat.completions.create({
  model: "google/gemini-4-carbon",
  max_tokens: 32768,
  stream: true,
  messages: [
    {
      role: "user",
      content:
        "Two workers retry the same payment after a timeout, and some customers are charged twice. Explain the race in three sentences.",
    },
  ],
});

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? "");
}

An image, sent inline as a base64 data URL (Python, same client as above):

import base64

with open("architecture.png", "rb") as f:
    image_b64 = base64.b64encode(f.read()).decode()

response = client.chat.completions.create(
    model="google/gemini-4-carbon",
    max_tokens=32768,
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "List the single points of failure in this architecture diagram, numbered."},
            {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{image_b64}"}},
        ],
    }],
)
print(response.choices[0].message.content)

Rules that hold for every request:

• Set max_tokens on every request. Leave room above the longest answer you expect; the samples above set 32768, the minimum Google recommends for an explicit output limit on this model (What's new in Gemini 4 Carbon, checked 2026-11-11). On Router One's Chat Completions, max_completion_tokens is accepted as the same cap (if both are sent, max_completion_tokens wins). Per Google's thinking guide (checked 2026-11-11), thinking tokens count toward the output cap, so leave headroom.
• Leave out sampling parameters. Per the migration checklist on Google's What's new page (checked 2026-11-11), remove temperature, top_p and top_k; on Google's own Gemini API they are ignored if provided. Put tone and format rules in the system message instead.
• Allow for long requests. Per the same page (checked 2026-11-11), hard prompts can take several minutes on Google's own API, so give your HTTP client a generous timeout and stream long answers with stream: true (streaming guide).
• Keep a healthy wallet balance. Before a request runs, the gateway reserves an estimate against your balance; when the balance cannot cover it, the gateway lowers max_tokens to fit, which can cut a long answer short. If it cannot cover even the request's initial reservation, the gateway returns HTTP 402 before the model runs.
• Send images inline. Pass images as base64 data URLs in image_url content parts, as in the example above; Router One serves no Files API.
• Test what your code depends on. Before you move a workload, run the parameters and request shapes your code sends (tool definitions, image inputs, output caps, any thinking setting) on your own prompts first.

5. Read the trace. Dashboard → Logs shows each call's model, tokens, wallet charge or plan quota used, latency and status (per-request observability). Retryable upstream failures are absorbed by automatic fallback across the candidate routes for the same model; a Gemini 4 Carbon request is never answered by a different model.

A request the gateway refuses before any model is called has no row in Dashboard → Logs; its status code and error message say why. A 401 means the key was not accepted, a 402 that the wallet balance or the key's maxSpend cannot cover the request's initial reservation, and a 404 that the path or the model id does not exist, so copy the id exactly as the catalog lists it. The error codes page covers the rest.

Which clients can call it

Any client that speaks Chat Completions and takes a custom base URL can call Gemini 4 Carbon: the official OpenAI SDKs (OpenAI SDK setup) and the chat apps and frameworks in the integrations directory that send /v1/chat/completions requests. Give the client the base URL https://api.router.one/v1 and a Router One key, and enter google/gemini-4-carbon as the model; the OpenAI-compatible API page explains the endpoint in general.

Claude Code cannot call Gemini 4 Carbon. Claude Code sends Anthropic Messages requests to /v1/messages, which Router One serves for Claude-family and DeepSeek ids only; for any other chat id the catalog lists, the gateway returns HTTP 400 before any model is called. Call Gemini 4 Carbon from a Chat Completions client. The API compatibility facts state the endpoint rule for each model family.

From mainland China

Requests reach api.router.one from mainland China without a VPN, on the same key and base URL. Top up with a card or Alipay through one hosted checkout, or with USDT/USDC on six chains (Tron, BSC, Ethereum, Polygon, Base, Arbitrum). No US credit card required. For the other Gemini models Router One lists, see the Gemini API in China page and the Gemini API in China guide.

FAQ

What is the model ID for Gemini 4 Carbon on Router One? The catalog id is google/gemini-4-carbon, as of the 2026-11-11 catalog; use it in code and client settings. The gateway also accepts Google's own id, gemini-4-carbon, for the same catalog entry on /v1/chat/completions. As of the 2026-11-11 plan response, no tier of the Pro, Max or Ultra plans lists it, so its calls bill per token to the wallet.

How much does the Gemini 4 Carbon API cost? On Router One, per token at the input and output rates posted on the Gemini 4 Carbon model page, with pricing up to 90% off official prices on select models; this guide prints no figures because they go stale. As of the 2026-11-11 plan response, it is in no plan tier, so every call bills the wallet. On Google's own Gemini API, per Google's pricing and What's new pages (checked 2026-11-11), the model is at an introductory price through November 30, 2026, and Google's standard pricing takes effect on December 1, 2026.

Is there a free tier for Gemini 4 Carbon? Not on Google's own Gemini API: per Google's pricing page (checked 2026-11-11), Gemini 4 Carbon is not available on Google's free tier. On Router One, every Gemini 4 Carbon call bills the wallet at the posted rates; as of the 2026-11-11 plan response, it is in no plan tier.

Do I need a Google API key to call Gemini 4 Carbon? Not through Router One: one Router One key (sk-...) on https://api.router.one/v1 calls Gemini 4 Carbon and the other catalog models, with no Google API key and no Google Cloud billing account.

What is the context window of Gemini 4 Carbon? Per Google's model page and What's new in Gemini 4 Carbon page (checked 2026-11-11), Gemini 4 Carbon has a context window of 1,048,576 tokens that input, thinking and output share — the unified 1M-token window Google describes — with a default maximum output of 65,536 tokens on Google's own Gemini API. The Router One model page shows the live catalog entry for the id.

Which thinking levels does Gemini 4 Carbon support? On Google's own Gemini API, per Google (checked 2026-11-11): low, medium and high, with medium as the default; reasoning cannot be turned off for Gemini 4 models, and, per Google's OpenAI compatibility guide, reasoning_effort sets the level on Google's OpenAI-compatible endpoint. The samples in this guide do not set reasoning_effort; test any setting your code sends on your own prompts first.

Should I use Gemini 4 Carbon or Gemini 4 Argon? Per Google's Gemini models page and What's new page (checked 2026-11-11), Gemini 4 Carbon is Google's strongest reasoning model and spends more thinking tokens per turn, while Gemini 4 Argon is Google's efficient Gemini 4 frontier reasoning model, with faster responses and lower token spend per turn. On Router One both are served on /v1/chat/completions, so switching is a model-string change. As of the 2026-11-11 plan response, Gemini 4 Argon is in the Premium models tier of the Pro, Max and Ultra plans. As of the 2026-11-11 plan response, no tier lists Gemini 4 Carbon.

Is Gemini 4 Carbon included in Router One subscription plans? No, as of the 2026-11-11 plan response: no tier of the Pro, Max or Ultra plans lists it, so its calls bill the wallet per token at the posted rates, plan or no plan. Gemini 4 Argon, a separate catalog entry, is in the Premium models tier of those plans as of the 2026-11-11 plan response. The pricing page lists each plan's allowance.

Can I use Gemini 4 Carbon in Claude Code? No. Claude Code sends Anthropic Messages requests to /v1/messages, which Router One serves for Claude-family and DeepSeek ids only, so a Gemini 4 Carbon request there returns HTTP 400 before any model is called. Call Gemini 4 Carbon from a Chat Completions client.

Can I call Gemini 4 Carbon from mainland China? Yes: api.router.one is reachable from mainland China without a VPN, on the same key and base URL. Top up with a card or Alipay through one hosted checkout, or with USDT/USDC on six chains (Tron, BSC, Ethereum, Polygon, Base, Arbitrum). No US credit card required.

Next steps

• Open the Gemini 4 Carbon model page for the live rates and the endpoint.
• Put it beside Gemini 4 Argon on the Gemini 4 Carbon vs Gemini 4 Argon comparison page, and read the Gemini 4 Argon API guide.
• Check each plan's allowance and the models each tier lists on the pricing page.
• Calling other Gemini models from mainland China? See the Gemini API in China page and the Gemini API in China guide.
• New to OpenAI-compatible endpoints? The OpenAI-compatible API page covers base URLs, keys and SDKs.
• For Gemini 3.8 Flash, read the Gemini 3.8 Flash API guide.