Modellix

Gemini 4 Carbon: Release Date, Pricing & How to Use It

Gemini 4 Carbon went GA on November 10, 2026. Here's what it does, what it costs before and after December 1, and how to try it on Modellix.

gemini 4 carbon is live on modellix
On this page
On this page
Summary
Google made Gemini 4 Carbon generally available on November 10, 2026. It's Google's strongest model so far, built for long agentic tasks, hard debugging, and security work. It costs $2 / $10 per million input / output tokens until November 30, then doubles. As of November 16, 2026, it's live on Modellix as google/gemini-4-carbon.

Google announced Gemini 4 Carbon on October 28 and opened it to every paying developer just 13 days later. That's a fast turnaround for a model Google calls its strongest yet.

The pitch is pretty clear. Carbon is meant for the work that falls apart halfway through: long agent runs, bugs buried deep in a big codebase, and security work. You can now call it on Modellix too, so we gave it a small bug to fix and checked the answer ourselves.

Here's what's new, what it costs (and why December 1 matters), how our quick test went, and a request you can copy.

Gemini 4 Carbon at a Glance

Item Details
Provider Google
Generally available November 10, 2026
Good at Long-running agent work, multi-step engineering, cyber defense
Inputs and outputs Text, images, video, audio, and PDFs in; text out
Context window / max output 1,048,576 tokens / 65,536 tokens by default
Thinking levels Low, medium, high
Google API model ID gemini-4-carbon
Google price until Nov 30 $2 per million input tokens / $10 per million output tokens
Google price from Dec 1 $4 / $20 per million tokens 2x price
Modellix model ID google/gemini-4-carbon
Modellix price $2 / $10 per million tokens, same as Google's current rate
Last checked November 16, 2026
Note: Specs and Google pricing come from Google's Gemini 4 Carbon model page and Gemini API pricing page, checked November 16, 2026. Thinking tokens are billed as output. Output can go past 65,536 tokens only with long decode turned on.
aistudio gemini 4 carbon
Source: Gemini API Model Documentation

How We Got Here: A Quick Timeline

Date What happened
October 28, 2026 Google announces Gemini 4 Carbon. Early access goes to a small group of cyber defenders in its Fairwind Program
November 10, 2026 General availability: the Gemini API and Gemini for business for paying developers and enterprises, plus a rollout to Google AI Ultra subscribers in the Gemini app
November 12, 2026 Gemini 4 Carbon goes live on Modellix
December 1, 2026 Google's introductory price ends, and rates double to $4 / $20
Note: Dates come from Google's October 28 announcement and November 10 availability post. Cyber-defense access still goes through the Fairwind Program, on a rolling basis.

What's New in Gemini 4 Carbon

Google puts Carbon above Gemini 4 Argon in the lineup. Here's what stands out:

  • Long, multi-step work. Google's main pitch is sustained agentic workflows, where the model keeps going over many steps without losing track. Think root-causing and patching a subtle bug across a large codebase, or digging into low-level systems code.
  • Security work. Carbon is built to find, validate, and fix vulnerabilities. That's why the first people to get it were cyber defenders.
  • More thinking, more tokens. You can pick a low, medium, or high thinking level. Google says Carbon uses more tokens than Argon, and thinking tokens count as output. So a short answer can still come with a surprisingly big output bill. Keep high for the hard problems.
  • Text out only. Carbon reads text, images, video, audio, and PDFs, but it only writes text. No image or audio generation, no Live API, and no explicit context caching. It's also not on the Gemini API free tier.

Benchmark Snapshot

These are vendor-reported numbers from Google's announcement. We haven't reproduced any of them.

Benchmark Gemini 4 Carbon What it measures Worth knowing
Artificial Analysis Intelligence Index v5.0 59 (#1) Overall intelligence across many evals Third-party index, as cited by Google
SEC-bench Pro 88.2% Finding vulnerabilities Evaluated without guardrails
ExploitGym 54.8% Exploiting vulnerabilities Evaluated without guardrails
CWE-bench 71.0% Patching vulnerabilities Evaluated without guardrails
WebDev Arena 1768 Elo (#4) Building web apps, judged by people +89 Elo, per Google
gemini 4 carbon evals table original
Note: Scores from Google's October 28 announcement, checked November 16, 2026. All results are vendor-reported, not Modellix tests.

One thing to keep in mind about the cyber scores: Google ran those evals without guardrails. They show what the raw model can do, not necessarily what you'll see with safety filters switched on.

What We Saw in a Quick Test

Since Carbon's big selling point is fixing tricky bugs, we gave it a small one. The function below is supposed to merge overlapping ranges. It works most of the time, but it cuts off part of the range when one interval sits fully inside another. We sent the function and the failing test to google/gemini-4-carbon with this prompt:

text
The test test_contained_intervals fails for the Python function below. Find the root cause and fix it. Keep the documented behavior: touching intervals merge, the result is sorted by start, and the input list is not modified. Reply with one sentence explaining the root cause, then the full corrected function in a single Python code block. Do not claim to have run the tests.

def merge_intervals(intervals):
    """Merge overlapping closed intervals.

    Touching intervals such as (1, 3) and (3, 5) merge. The result is sorted
    by start, and the input list is not modified.
    """
    if not intervals:
        return []
    ordered = sorted(intervals)
    merged = [list(ordered[0])]
    for start, end in ordered[1:]:
        last = merged[-1]
        if start <= last[1]:
            last[1] = end
        else:
            merged.append([start, end])
    return [tuple(pair) for pair in merged]

def test_contained_intervals():
    assert merge_intervals([(1, 10), (2, 3), (4, 5)]) == [(1, 10)]

Here's what Carbon sent back:

The function unconditionally overwrites last[1] with end upon finding an overlap, which shrinks the merged interval when a subsequent interval is completely contained within it rather than keeping the larger end value.

python
def merge_intervals(intervals):
    """Merge overlapping closed intervals.

    Touching intervals such as (1, 3) and (3, 5) merge. The result is sorted
    by start, and the input list is not modified.
    """
    if not intervals:
        return []
    ordered = sorted(intervals)
    merged = [list(ordered[0])]
    for start, end in ordered[1:]:
        last = merged[-1]
        if start <= last[1]:
            last[1] = max(last[1], end)
        else:
            merged.append([start, end])
    return [tuple(pair) for pair in merged]

We didn't just take its word for it. We ran the patched function against our five unit tests, then threw 3,000 random cases at it and compared every result with a brute-force answer.

Measurement What we saw
Request One non-streaming request, HTTP 200
Found the real cause? Yes. It spotted that the end gets overwritten instead of extended
Unit tests passed 5 / 5
Random cases passed 3,000 / 3,000
Time to answer 90.2 seconds
Input / output tokens 297 / 284 (incl. 78 thinking)
Cost About $0.003
Note: One request on November 16, 2026, default thinking level, billed at Modellix's rate. Time includes connection and transfer. This is a quick sanity check on one small function. It doesn't tell you much about long agent runs, reliability across many runs, or Carbon's security skills.

Our quick take: Carbon found the real bug in one sentence and changed exactly one line, and the fix held up against everything we threw at it. The slow part was the wait: about 90 seconds for a short answer. Carbon had only been out for a few days, so launch-week load may have played a part. Time it on your own tasks before you put it in anything interactive.

What This Means for Modellix Users

You can call Carbon with the Modellix API key you already have. Just use the fixed model ID google/gemini-4-carbon. Stick with the fixed ID when you want results you can repeat. There's no latest alias for Carbon yet: ~google/gemini-pro-latest still points to google/gemini-3.1-pro.

Already using google/gemini-4-argon? Switching is basically a one-line change. Google gave Carbon the same introductory price as Argon, but Carbon tends to use more tokens. So before you move everything over, compare the cost per task, not just the price per token.

To get a feel for the bill: a job with 1M input tokens and 200K output tokens costs $4 at Google's introductory rate and $8 from December 1. On Modellix, at today's $2 / $10 rate, the same job costs $4.

One API for every model

Run Gemini 4 Carbon next to every other model on one key

One key, every model

GPT-6 Astra, Sol, and Luna, plus Claude, Gemini, Grok, DeepSeek, and Qwen. Switching models is a change to the model string.

Drop-in for your client

OpenAI Chat Completions and Responses, plus Anthropic Messages. Works with the OpenAI and Anthropic SDKs, Codex, Claude Code, Cursor, and OpenCode.

Below list price

GPT-6 and Claude models bill 10% under the vendor list price on the current rate card. Pay as you go, no subscription.

Cost per request, in the logs

Every call is logged with tokens and cost, and can be tagged by end user, so you can see what an escalation actually cost.

Pin or follow

Pin a fixed model ID for evaluations, or use a -latest alias that moves to the newest release.

Media on the same key

The same account also calls image, video, and audio generation models through the Modellix media API.

How to Try Gemini 4 Carbon on Modellix

Google models on Modellix use the OpenAI-compatible Chat Completions endpoint, https://llm.modellix.ai/v1/chat/completions (the Responses API works too; the Anthropic Messages API doesn't). Set your MODELLIX_API_KEY environment variable, and you can send the same request we tested:

bash
curl -sS https://llm.modellix.ai/v1/chat/completions \
  -H "Authorization: Bearer $MODELLIX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemini-4-carbon",
    "max_tokens": 16384,
    "messages": [{"role": "user", "content": "<paste the prompt above>"}]
  }'
Note: If you get an HTTP 402 error with type `insufficient_quota`, your account balance is too low. Top up and send the request again.

FAQ

What's the Gemini 4 Carbon model ID?

It's gemini-4-carbon on Google's Gemini API and google/gemini-4-carbon on Modellix.

How much does Gemini 4 Carbon cost?

Google charges $2 per million input tokens and $10 per million output tokens through November 30, 2026. From December 1, that goes up to $4 and $20. Thinking tokens count as output. On Modellix, it's $2 / $10 right now, the same as Google's current rate.

Can Gemini 4 Carbon generate images or audio?

No. It can read text, images, video, audio, and PDFs, but it only writes text.

Is there a free way to try Gemini 4 Carbon?

Not on the Gemini API, since Carbon isn't on the free tier. On the consumer side, it's rolling out to Google AI Ultra subscribers in the Gemini app.

Start Building with Gemini 4 Carbon

The best way to size up Carbon is to hand it a bug you already understand and see if it finds the real cause, not just a quick patch. Check current pricing in the Modellix model catalog, grab your API key, and try the fixed model ID on your own code.

Modellix Team
Written byModellix TeamOfficial