DEV Community

Cover image for Gemini 4 Carbon Explained: Benchmarks, Pricing, and Who Can Use It Today
Muhammad Sheraz
Muhammad Sheraz

Posted on

Gemini 4 Carbon Explained: Benchmarks, Pricing, and Who Can Use It Today

#ai #gemini #cybersecurity #programming

Gemini 4 Carbon Explained: Benchmarks, Pricing, and Who Can Use It Today

Google shipped Gemini 4 Argon to the public on October 13. On October 28, it announced Gemini 4 Carbon, a model it calls its strongest yet. That is two weeks between frontier releases.

If you build on LLM APIs, that cadence matters as much as any single benchmark number. Here is what Google announced, what independent sources say, what is still unknown, and how to get your evals ready before the API opens.

📅 Status as of October 29, 2026: Carbon is announced but not generally available. I have not had access to it, so this post covers published results and reports, not hands-on testing.

🤔 What is Gemini 4 Carbon?

Carbon is the next model in Google's Gemini 4 family, following Argon. Google positions it around two things: cybersecurity and long-horizon work, meaning agentic tasks that run for a long time and span many steps.

The name first surfaced on October 10, when Business Insider reported that Google employees were testing a checkpoint called Carbon on Jetski, Google's internal coding platform. One employee said its coding felt comparable to Claude Opus 5.5, while early Argon builds felt closer to Opus 5. That was an anecdote, not a benchmark.

📊 The numbers Google published

Area Result for Carbon
Artificial Analysis Intelligence Index v5.0 59, ranked #1
SEC-bench Pro (finding vulnerabilities) 88.2%
ExploitGym (validating exploitability) 54.8%
CWE-bench (producing working fixes) 71.0%
WebDev Arena +89 Elo, #4 at 1769

Google says the cyber scores were measured without guardrails, meaning the raw model, not the safeguarded product.

Independent signals line up with the headline:

  • Artificial Analysis ranks Carbon #1 overall, but second to Claude Opus 5.5 on its Coding Agent Index.
  • Arena reported Carbon at #1 in Text Arena, #2 in Agent Arena, and #4 in the WebDev category.

So "best model" depends on the leaderboard. For general intelligence, Carbon leads. For agentic coding, Opus 5.5 still holds the top spot on that index.

🛡️ Safety and reliability claims

Google also published reliability numbers, which matter if you plan to run an agent unattended:

  • Terminal-Bench 4.0: 0.0% successful cheating, with cheating attempts at 0.4%.
  • Internal coding work: high-severity misalignment alerts down by almost 80%.
  • Honesty: roughly half as many unsupported claims as Argon.
  • Prompt injection: attack success on Gray Swan's indirect prompt injection benchmark fell from Argon's 0.7% to 0.0%.

These are Google's own internal results. Treat them as a hypothesis to test on your workloads.

🔒 Who can use it right now?

Almost nobody. Access is staged:

  1. This week: a subset of cyber defenders.
  2. Next: additional trusted partners, added gradually through Google DeepMind's Fairwind Program.
  3. Later: developers, businesses, and consumers. Google says "soon" and has given no date.

Google ties the staged rollout to its participation in the US government's voluntary pre-release testing process, with wider access held back while safeguards are tested on trusted defenders.

💰 Pricing

Reporting says Carbon will launch at Argon's introductory price:

Input (per 1M tokens) Output (per 1M tokens)
Argon, intro price (through Nov 30) $2 $10
Argon, from Dec 1 $4 $20
Carbon (reported launch price) $2 $10

Cached input is reported at 95% less than standard input. Two caveats:

  • Google has not said how long Carbon's launch price lasts. Argon's promo ends November 30, so budget at the higher rate until Google confirms.
  • Context window and rate limits for Carbon have not been published.

⚖️ Carbon vs Argon

Argon Carbon
Availability Generally available since Oct 13 (gemini-4-argon) Defenders first, no public date
Positioning Advanced reasoning, knowledge work, coding Cybersecurity and long-horizon engineering
Prompt injection (Gray Swan) 0.7% attack success 0.0%
Honesty Baseline ~2x fewer unsupported claims (Google internal)
Speed Faster Pre-release testers reported 2–4x slower

The speed row comes from pre-announcement testers on hand-picked prompts, and nobody could be sure which checkpoint they were served. The final model may differ, so don't plan around "2–4x slower" yet.

🧪 What pre-launch testers said

Before the announcement, early-access chatter described Carbon as stronger at 3D generation, "less lazy" about finishing long tasks, and still weak at frontend and UI design. That matches the pitch of a model tuned for sustained, verifiable engineering work, but it is anecdotal. Wait for measured results.

🧰 How to prepare before the API opens

You can't call Carbon today, but you can make switching cheap.

1. Make the model name a config value, not a constant. Argon's identifier is public. Carbon's isn't, so don't hard-code a guess.

import os
import time
from google import genai

client = genai.Client()  # reads GEMINI_API_KEY from the environment

MODEL = os.environ.get("GEMINI_MODEL", "gemini-4-argon")

# Prices per 1M tokens. Update when Google publishes real numbers.
PRICE_IN = float(os.environ.get("PRICE_IN", "4"))    # budget at the Dec rate
PRICE_OUT = float(os.environ.get("PRICE_OUT", "20"))


def run_task(prompt: str) -> dict:
    start = time.perf_counter()
    resp = client.models.generate_content(model=MODEL, contents=prompt)
    elapsed = time.perf_counter() - start

    usage = resp.usage_metadata
    tokens_in = usage.prompt_token_count
    tokens_out = usage.candidates_token_count
    cost = tokens_in / 1e6 * PRICE_IN + tokens_out / 1e6 * PRICE_OUT

    return {
        "model": MODEL,
        "seconds": round(elapsed, 1),
        "tokens_in": tokens_in,
        "tokens_out": tokens_out,
        "cost_usd": round(cost, 4),
        "text": resp.text,
    }
Enter fullscreen mode Exit fullscreen mode

2. Check which models your key can see before assuming anything:

for m in client.models.list():
    print(m.name)
Enter fullscreen mode Exit fullscreen mode

3. Build a small regression suite from your own work. Public benchmarks measure someone else's tasks. Pull 20 to 50 real examples: tickets you closed, bugs you fixed, reviews you did. Store the prompt, the expected behavior, and a pass/fail check that runs code, not vibes.

4. Track three numbers per model: pass rate, wall-clock time, and cost per successful task. If a model is slower but finishes tasks in one pass, cost per success can still win.

5. Pin versions in production. With frequent checkpoints, silent behavior changes are a real risk. Pin, then re-run your suite on every new release.

🔐 A defensive test worth running

Since Google is pitching Carbon on find-and-fix security work, build a patch evaluation on your own code. Take a known vulnerable pattern from your repo's history, like this one:

// Vulnerable: string-concatenated SQL
app.get("/users", async (req, res) => {
  const rows = await db.query(
    `SELECT * FROM users WHERE email = '${req.query.email}'`
  );
  res.json(rows);
});
Enter fullscreen mode Exit fullscreen mode

Ask each model for a patch, then apply it and run your tests plus a regression test for the original issue. Score on whether the patch actually fixes the bug without breaking anything else. That is closer to what CWE-bench measures than a chat-window impression.

🎯 Who should care

  • Security teams and platform engineers: the Fairwind route is the one to watch if you work on vulnerability management or patching.
  • Teams running long agent loops: the reliability claims (less cheating, fewer unsupported claims, better injection resistance) are the interesting part, once you can verify them.
  • Everyone else: stay on Argon or your current model, keep your evals ready, and revisit when the API opens.

⚠️ What's still unknown

  • Public release date for developers and consumers
  • The API model identifier
  • Context window and rate limits
  • Whether launch pricing persists past Argon's November 30 promo
  • Real-world latency compared with Argon
  • Independent results on tasks outside Google's chosen benchmarks

🏁 Conclusion

Gemini 4 Carbon puts Google back at the top of at least one major leaderboard, with strong published results on vulnerability discovery, exploitation, and patching. Access is limited, key specs are missing, and Opus 5.5 still leads the agentic coding index. The practical takeaway is that frontier models now ship in weeks, so a portable eval suite and pinned versions are worth more than any single benchmark score.

💬 Your turn

Are you in a position to get Carbon access through Fairwind, or will you wait for the public API? Tell me which workloads you'd test first in the comments, and follow for an update when the model opens up. I'll publish real numbers on speed and cost per task once I can run it.

Sources

Top comments (0)