API Market / AI News / Gemini 4 Carbon: what we know about Google's new frontier model

Gemini 4 Carbon: what we know about Google's new frontier model

Published: November 4, 2026 · Official

On October 28, 2026, Google DeepMind announced Gemini 4 Carbon, the frontier model that follows Gemini 4 Argon. Google claims state-of-the-art cybersecurity and long-horizon agentic work, with launch pricing matching Argon. General access is not available yet: the model is being phased out to trusted partners.

What Google officially announced

In its October 28 blog post, Google DeepMind calls Gemini 4 Carbon its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. The announcement is signed by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google.

The model builds on Gemini 4 Argon, launched roughly two weeks earlier. Google says it is pushing the frontier faster while strengthening safeguards in tandem, and is rolling Carbon out deliberately: first to a subset of cyber defenders, then gradually to trusted partners through the Fairwind Program.

Benchmarks: first place by a single point

According to Artificial Analysis, Gemini 4 Carbon scores 59 on the updated Intelligence Index v5.0, one point ahead of Claude Opus 5.5 (58) and two ahead of GPT-6 Astra (57). Google cites the same figure in its post.

On AA-Omniscience, which rewards correct answers and penalises hallucinations, Carbon scores 54 against 48 for Opus 5.5 and 46 for Gemini 4 Argon. Its hallucination rate is 11%, the second lowest after Argon at 4%, while Opus 5.5 and GPT-6 Astra are both listed at 50%.

Tom's Hardware puts it more cautiously: Carbon narrowly beats Claude Opus 5.5 and is claimed to hallucinate least. Artificial Analysis ran the tests; no one else has repeated them yet.

Where the model trails

Carbon is not first on coding agents: on the AA Coding Agent Index with the Antigravity CLI it scores 59, three points behind Claude Code with Opus 5.5 at 62. Opus 5.5 also leads Terminal-Bench 4.0 and FrontierCode.

On AA-Briefcase, which covers knowledge work, Carbon scores 1609, up 126 from Argon but behind Grok 4.7 (1638), Muse Wonder (1659) and Kimi K3.1 Pro (1678). It ranks fourth on WebDev Arena with 1768 points.

9to5Google, citing SemiAnalysis' GDPx benchmark, notes that removing the cybersecurity, biology and alignment domains shuffles the leaderboard: GPT-6 Astra, GPT-6.1 Sol and Claude Sonnet 5.5 move ahead of Carbon. The same report puts cost at about $16 per task for Carbon against roughly $51 for Opus 5.5.

Price and availability

Launch pricing matches Argon: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off the input price. According to tbreak, standard pricing is $4 and $20, with a 50% launch discount and no confirmed end date.

The context window is 1 million tokens, and the model accepts text, images, video and speech. Artificial Analysis estimates $7.76 per index task, about 23% less than Opus 5.5.

There is no public release date. 9to5Google reports the model is listed as coming soon for most customers, and Google has not said when developers, businesses and consumers will get access.

Cybersecurity as the main pitch

Google targets the full defensive lifecycle: finding vulnerabilities, validating exploitability and patching. By the company's own figures, Carbon reaches 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench.

Internal tests show more than a 2x uplift in generating secure code over Argon. The Chrome team reported Carbon performed more than 10x better than another lab's large frontier model when generating high-quality patches. On Terminal-Bench 4.0, successful cheating dropped to 0.0% with attempts down to 0.4%.

One caveat: these are Google's own measurements. No independent confirmation was available at the time of writing.

What to use right now

Carbon is limited to select partners for now, so for day-to-day work it makes sense to look at models that are already open. API Market's catalog includes Gemini 4 Argon, Google's senior model with a long context that Carbon builds on, and the fast Gemini 3.8 Flash for high-volume requests.

Where Carbon trails on benchmarks, Claude Opus 5.5 remains an option: Anthropic's flagship for complex code and analysis, available in the same catalog. You can compare models on your own tasks through a single API Market key, with payment in rubles and no VPN. Keep an eye on the API Market model catalog for new arrivals.

Models you can use right now

ModelAPI Market price
Gemini 4 Argon
gemini-4-argon		input $0.35, output $1.75 per 1M tokens
Claude Opus 5.5
claude-opus-5-5		input $0.90, output $4.50 per 1M tokens
Gemini 3.8 Flash
gemini-3.8-flash		input $0.07, output $0.33 per 1M tokens

Try these models on API Market

One API key for GPT, Claude, Gemini, DeepSeek, Grok and more. OpenAI-compatible endpoint, pay as you go, top-ups from about $0.15.

Get an API keyAll models

Connect in a minute

from openai import OpenAI

client = OpenAI(
    base_url="https://api-market1.shop/v1",
    api_key="sk-your-key",
)

reply = client.chat.completions.create(
    model="gemini-4-argon",
    messages=[{"role": "user", "content": "Hi! Tell me about yourself."}],
)
print(reply.choices[0].message.content)

FAQ

When will Gemini 4 Carbon be available to everyone?

No date has been given. Google is phasing access, starting with cyber defenders and Fairwind Program partners before opening it more widely. The company has not announced a general launch timeline.

Does Carbon really beat Claude Opus 5.5?

On the Artificial Analysis index, yes, but narrowly: 59 versus 58. On coding agents, Opus 5.5 still leads 62 to 59.

How much does Gemini 4 Carbon cost?

At launch, $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. According to tbreak, standard pricing is double that, and the launch discount has no confirmed end date.

Does Carbon hallucinate less?

Artificial Analysis puts its hallucination rate at 11%, below Opus 5.5 and GPT-6 Astra at 50% each, but above Gemini 4 Argon at 4%.

What is the context limit?

1 million tokens. The model accepts text, images, video and speech as input.

Sources

Русская версия