Gemini 4 Carbon API Access Guide (2026): Who Can Use It, $2/$10 Pricing, and What to Build on Today

On October 28, 2026, Google announced Gemini 4 Carbon, its new top model, and in the same post made clear that most developers can't have it yet. Carbon goes first to "a subset of cyber defenders," with more trusted partners added gradually through Google's Fairwind Program. There's no general release date. Ars Technica's headline summed it up: a frontier model that you can't use.

So there's no public model ID to paste into a curl command. Instead, here's what Google has published, what the pricing does to a real budget, and how to wire your code so Carbon's opening day is a config change.

Key Takeaways

What Google Actually Announced

Google's post, written by Koray Kavukcuoglu, calls Carbon "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." It follows Gemini 4 Argon by two weeks. Google also says it is taking part in the U.S. government's voluntary process for pre-release model testing, and that it wants to keep gathering feedback and iterating on guardrails before opening Carbon to developers, businesses, and consumers "as soon as possible."

That "as soon as possible" is the whole timeline. No month, no quarter. The numbers Google put up:

• Artificial Analysis Intelligence Index v5.0: 59, ranked first. Google calls v5.0 "recently updated."
• SEC-bench Pro (finding vulnerabilities): 88.2%.
• ExploitGym (exploitation): 54.8%.
• CWE-bench (patching): 71.0%, on Collinear's benchmark.
• Gray Swan indirect prompt injection: 0.0% attack success rate, down from 0.7% for Argon.

Two caveats. Google says the security benchmarks ran "without guardrails," so the model defenders get may behave differently. And these are Google-reported figures, not independent replications.

Gemini 4 Carbon API Pricing

All prices are USD per 1 million tokens. Carbon's come from Google's October 28 announcement, Argon's from Google's Gemini API pricing page (read October 30, 2026). "Derived" means my arithmetic on Google's 95%-off statement, not a published rate.

Argon got a hard end date (November 30). Carbon got only a footnote saying $4/$20 applies "after the introductory period expires." Budget Carbon at $4/$20 and treat $2/$10 as a bonus.

Carbon vs the Models You Can Call Today

Here's how Carbon stacks up against the models that are actually reachable by an API key this week.

My read: unless defensive security is your job, Carbon's absence costs you nothing this week. Argon shipped two weeks earlier at the same price. The real risk is designing around Carbon's benchmark halo before there's an endpoint to test.

What the Price Means on a Real Workload

Take a code-review agent that sends 200 million input tokens and generates 20 million output tokens per month. These are my example numbers, not Google's.

• Carbon at $2/$10, no caching: $400 input plus $200 output, so $600 per month.
• Carbon at $4/$20, no caching: $800 plus $400, so $1,200.
• Carbon at $4/$20, same 60% cached: 120M at $0.20 is $24, 80M at $4 is $320, output is $400. Total $744.
• GPT-6.1 Terra, Standard: $200 input plus $100 output, so $300.

The doubling is the story, and caching takes the sting out of it: $1,200 falls to $744 once most of the repeated context is cached. Run your own token counts through the API cost calculator, and measure real prompt sizes with the token counter before you trust any estimate.

Wiring Your Code for the Day Carbon Opens

I haven't made a Carbon call, because nobody outside the early cohort can. What I can show is the pattern: keep the model ID in configuration, default to Argon, and fall back automatically. The calls below use the Interactions API shown in Google's own Gemini 4 Argon quickstart.

curl "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H 'Content-Type: application/json' \
  -X POST \
  -d "{
    \"model\": \"${GEMINI_MODEL:-gemini-4-argon}\",
    \"input\": \"Review this diff for injection bugs: ...\"
  }"

import os
from google import genai

client = genai.Client()

# Point GEMINI_MODEL at Carbon's ID once Google publishes it.
PRIMARY = os.environ.get("GEMINI_MODEL", "gemini-4-argon")
FALLBACK = "gemini-4-argon"

def ask(prompt: str) -> str:
    for model in dict.fromkeys([PRIMARY, FALLBACK]):  # de-dupes if they match
        try:
            interaction = client.interactions.create(model=model, input=prompt)
            print(f"answered by {model}")
            return interaction.output_text
        except Exception as err:  # narrow to your SDK's API error type
            print(f"{model} failed: {err}")
    raise RuntimeError("all models failed")

print(ask("Review this diff for injection bugs: ..."))

import { GoogleGenAI } from "@google/genai";

const client = new GoogleGenAI({});
const models = [process.env.GEMINI_MODEL ?? "gemini-4-argon", "gemini-4-argon"];

async function ask(input) {
  for (const model of new Set(models)) {
    try {
      const interaction = await client.interactions.create({ model, input });
      console.log(`answered by ${model}`);
      return interaction.output_text;
    } catch (err) {
      console.error(`${model} failed:`, err.message);
    }
  }
  throw new Error("all models failed");
}

console.log(await ask("Review this diff for injection bugs: ..."));

Log which model answered every request, and run Carbon against Argon on your own eval set before flipping the default. A SEC-bench Pro lead says little about your codebase. To compare vendors behind one OpenAI-compatible endpoint while you wait, a gateway like KissAPI lets you swap the model string without touching client code.

What to Do Now

If you do defensive security work: Google's post describes no application process for the Fairwind Program, so ask your Google Cloud contact.

If you ship general product code: use Argon, and lock in your numbers before December 1, 2026. Test GPT-6.1 Terra or Claude Sonnet 5.5 alongside it; our Gemini 4 Argon pricing guide and GPT-6.1 Terra guide have the details. For long agent loops, assume $4/$20, turn on caching, and keep the fallback layer above.

Frequently Asked Questions

Can developers use the Gemini 4 Carbon API today?

No. As of October 30, 2026, Google has made Gemini 4 Carbon available only to a subset of cyber defenders, with additional trusted partners added gradually through the Fairwind Program. Google says it will bring Carbon to developers, businesses, and consumers "as soon as possible" but has not announced a date or an API model ID.

How much will Gemini 4 Carbon cost?

Google says Gemini 4 Carbon will launch at an introductory price of $2 per 1 million input tokens and $10 per 1 million output tokens, with cached input tokens priced 95% below the input price. After the introductory period, the price becomes $4 per 1 million input tokens and $20 per 1 million output tokens. Google's October 28, 2026 announcement does not state when the introductory period ends.

What should I use while waiting for Gemini 4 Carbon?

Gemini 4 Argon (model ID gemini-4-argon) is generally available in the Gemini API and costs the same $2 per 1M input and $10 per 1M output tokens through November 30, 2026, with a 1,048,576-token context window. Its price rises to $4 and $20 on December 1, 2026. GPT-6.1 Terra at $1 and $5, and Claude Sonnet 5.5 at $2 and $10, are also callable today.

Sources: Google, "Gemini 4 Carbon: our next frontier model," blog.google, October 28, 2026; Google AI for Developers Gemini API changelog, latest-model page, and Argon pricing (read October 30, 2026); Ars Technica and Reuters launch coverage, October 28, 2026. Benchmarks are Google-reported.