Gemini 4 Carbon: Google’s Top Model, Who Can Use It and What It Costs

1. The spec sheet: what Google published, and what it didn’t

Google’s launch post is long on benchmarks and short on specs. It gives a price and a rollout plan, but no model ID, context window, modality list or parameter count, and the URL pattern of Argon’s model card returns a 404 for Carbon. Where Google is silent, the box below says so and falls back to third-party listings, labeled as such.

The safety report linked from the post is Argon’s (October 2026), and I searched its text: Carbon isn’t mentioned. It says Google can’t rule out the unmitigated Argon reaching the Cyber Uplift Level 1 Critical Capability Level, which is why Argon shipped with cyber mitigations. For a successor sold on exploitation, that explains the staged rollout.

2. What improved vs Gemini 4 Argon

Google publishes its benchmark table as an image, so I take the rows from VentureBeat’s write-up, whose cyber numbers match the ones in Google’s post. By Google’s count, Carbon has the top score in 15 of 17 rows. All of them are vendor-reported, and Google says the cyber evaluations ran without guardrails. Against Argon: SEC-bench Pro 88.2% vs 83.4%, ExploitGym 54.8% vs 42.0%, CWE-bench 71.0% vs 68.0%, Agents’ Last Exam v1.1 77.4% vs 69.0%, GraphWalks 96.2% vs 84.2%, and FrontierSWE 2 61.2% vs 55.0%.

The independent check is Artificial Analysis. Its Carbon vs Argon comparison, which I read on November 3, has the Intelligence Index v5 going from 50 to 59 at high reasoning, Terminal-Bench 4.0 from 57% to 65% and FrontierCode 2.0 from 26% to 41%. If you compare with my Argon post, which quoted 53: version 5.0 changed the test mix, so old and new scores don’t mix.

The three changes Google is selling

3. Quickstart: build on Argon now, switch with one line

You can’t call Carbon today unless you’re in the first cohort: Google hasn’t published a model ID, and its post names no product surface for developers. So the code below runs on gemini-4-argon, with the model ID read from configuration. The second sample uses background execution, the documented way to run the long jobs Carbon is pitched at: the server runs the task, you get an interaction ID back right away, and you poll it without holding a connection open.

The snippets follow Google’s docs; I haven’t run them. You need a GEMINI_API_KEY from a paid-tier project, because Argon isn’t on the free tier. When Carbon opens, read its docs page for differences (thinking levels, output limits, caching) before you flip the variable.

# pip install -U google-genai
# export GEMINI_API_KEY="YOUR_API_KEY"  (paid-tier project: Argon has no free tier)
import os

from google import genai

# SWAP POINT: the only line that changes when Carbon opens.
# Set GEMINI_MODEL to the ID Google lists on its models page. Don't guess it.
MODEL = os.environ.get("GEMINI_MODEL", "gemini-4-argon")

client = genai.Client()

with open("payments/retry.py", encoding="utf-8") as f:
    source = f.read()

interaction = client.interactions.create(
    model=MODEL,
    input="Find race conditions in this retry logic and propose a minimal patch:\n\n" + source,
    generation_config={
        "thinking_level": "medium",  # low | medium (default) | high
        "max_output_tokens": 32768,  # thinking and the answer share this limit
    },
)

print(interaction.output_text)

import os
import time

from google import genai

MODEL = os.environ.get("GEMINI_MODEL", "gemini-4-argon")  # same swap point
BUDGET_SECONDS = 45 * 60  # your limit, not Google's

client = genai.Client()

with open("audit_bundle.txt", encoding="utf-8") as f:
    bundle = f.read()  # the files you want audited, concatenated

interaction = client.interactions.create(
    model=MODEL,
    input="Audit this code for injection and broken access control. For each "
    "finding give file, line, severity and a minimal patch.\n\n" + bundle,
    background=True,  # needs store=True, the default
    generation_config={
        "thinking_level": "high",
        "max_output_tokens": 131072,  # without it, background stops at 64K
    },
)
print("interaction id:", interaction.id)  # save it: you can poll or cancel later

started = time.monotonic()
while interaction.status in ("in_progress", "queued"):
    if time.monotonic() - started > BUDGET_SECONDS:
        client.interactions.cancel(id=interaction.id)
        raise TimeoutError(f"Cancelled {interaction.id} after {BUDGET_SECONDS} s")
    time.sleep(10)
    interaction = client.interactions.get(id=interaction.id)

if interaction.status != "completed":
    raise RuntimeError(f"Finished with status: {interaction.status}")

# Output over 64K arrives as several model_output steps: join them all.
text = "".join(
    content.text
    for step in interaction.steps
    if step.type == "model_output"
    for content in step.content
    if content.type == "text"
)
print(text)

# Lists the models your key can see and keeps the Gemini 4 names.
# Today this should print only Argon.
curl -s "https://generativelanguage.googleapis.com/v1beta/models?pageSize=1000" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  | grep -oE '"name": *"models/gemini-4[^"]*"'

4. What a long job costs, at intro and list prices

Google prices Carbon exactly like Argon, so Argon’s rows on the Gemini API pricing page are the best guide: $2 input, $0.10 cached input and $10 output per 1M tokens through November 30, then $4, $0.20 and $20, with no long-context surcharge listed. Carbon’s promo has no end date. Here is one agent run priced on paper; the assumptions are mine, so swap in numbers from your own logs.

Three things stand out. The promo halves every line, and Carbon’s has no end date, so budget at $4/$20 from day one. Caching is worth more than the promo: $9.87 without cache hits, $2.10 with them. And at list prices Carbon lands next to Opus 5.5 ($4.20 vs $4.45) and at about twice GPT-6.1 Sol, so Artificial Analysis’s cost per task ($7.76 Carbon, $10.04 Opus 5.5, $1.27 Sol) is the number to watch.

5. Access: who is in, and how to ask

I covered Fairwind’s eligibility rules in my Gemini 4 Argon guide; here is what is different for Carbon, and one thing that changed on the program page since then.

On Google Cloud, Fairwind already has a concrete shape. The Vertex AI page for Gemini 4 Argon lists a second model ID, gemini-4-argon-cyber-permissive, for vulnerability discovery, reverse engineering, exploit verification and patch generation. It sits behind an allowlist (requests go through your Google team or Fairwind), runs only in the global and us locations and doesn’t support computer use. A permissive Carbon could show up there the same way, but Google hasn’t said it will.

6. Honest limits

Plan around these before you promise Carbon to anyone.

Verdict

Gemini 4 Carbon looks like Google’s best model: nine points over Argon on the hardest version of Artificial Analysis’s index, top scores on vulnerability finding and exploitation, and an intro price no higher than Argon’s. For security teams and long autonomous engineering jobs, it’s the model to get in line for.

For everyone else it isn’t a decision yet, because you can’t call it. Build on Argon now, keep the model ID in configuration, run long tasks in background mode, and measure cost per verified task, so you can tell within a day whether Carbon earns its place when it opens.