Google just took the top spot in the AI race — and almost nobody can use the model yet. On October 28, 2026, Google DeepMind announced Gemini 4 Carbon, its second frontier model in a single month, and it immediately claimed first place on the independently run Artificial Analysis Intelligence Index v5.0 with a score of 59. The catch? Google is handing it to cyber defenders first, and there is no public launch date.
Here is what Carbon actually is, what the scores mean, what it costs, and when you might finally get your hands on it.
What Is Gemini 4 Carbon?
Gemini 4 Carbon is Google DeepMind's newest proprietary frontier AI model, announced by SVP Koray Kavukcuoglu on the Google blog on October 28, 2026. Google calls it "our strongest model with state-of-the-art cybersecurity and long horizon capabilities."
The timing is striking: Carbon arrived just two weeks after Gemini 4 Argon, which Google launched at the end of September and opened to paying users worldwide on October 13. Carbon is positioned as the stronger successor — built for much harder, long-horizon engineering problems. Google says its internal teams ran the model for days at a time on multi-step engineering tasks, framing it as a sustained-agent model: a system designed to run autonomously for days without losing the thread.
The practical pitch: an AI that finds software flaws, proves they are real, and patches them — then keeps working for days on the kind of long engineering tasks that break other models' focus.
Carbon vs Opus 5.5: The Benchmark Picture
Let's start with the headline number. On version 5.0 of the Artificial Analysis Intelligence Index — nine evaluations run independently by Artificial Analysis — Carbon scores 59, taking the #1 spot:
| Model | AA Intelligence Index v5.0 | Status |
|---|---|---|
| Gemini 4 Carbon (high) | 59 | Limited pre-release |
| Claude Opus 5.5 (max) | 58 | Available |
| GPT-6 Astra (max) | 57 | Available |
| Gemini 4 Argon | 50 | Available |
That is a nine-point jump over Argon and a one-point lead over Anthropic's Opus 5.5 — a thin margin, but a real one on independently run tests. On the separate Vals Index v2.2, Carbon sits at 70.36%, good for #2 overall.
The honest caveats, from independent analysts at Top5Apps and AI Insiders:
- Coding agents: On the Artificial Analysis Coding Agent Index, Carbon scores 59 — three points behind Claude Opus 5.5's 62. The headline index lead does not mean it wins everywhere.
- Agentic knowledge work: On AA-Briefcase, Carbon trails Kimi K3.1 Pro, Muse Wonder, and Grok 4.7. (See our Kimi K3.1 Pro and Muse Wonder explainers.)
- Web development: Carbon gained 89 Elo on WebDev Arena to 1,769 — still only fourth place.
- The security scores are Google's own numbers and, as The Decoder notes, cannot be independently verified yet.
A one-point overall lead with losses on coding agents and knowledge-work tasks: Carbon is the best composite model on the board, but not the best at every job.
Cybersecurity First: The Defense Angle
What makes this launch genuinely unusual is the framing. Google ran the security benchmarks without safety guardrails and reported three headline scores:
- 88.2% on SEC-bench Pro — finding real-world vulnerabilities
- 54.8% on ExploitGym — turning a disclosed flaw into a working exploit to validate risk
- 71.0% on CWE-bench — patching flaws to the root cause
Google calls each a state-of-the-art result. The unguarded-evaluation detail matters for two reasons: it explains why access is restricted (this is a model that can find and exploit real bugs), and it means defenders using the safeguarded version will be working with a different system — Google does not say how much the guardrails change the scores.
For context, Google's own Wiz cloud-security unit reports Carbon leading on web penetration testing benchmarks. And on Terminal-Bench 4.0, Google says successful cheating fell to 0.0%, with attempts down to 0.4% — accuracy still improving while gaming of the test collapses.
Gemini 4 Carbon Pricing
Google named the price in the launch post, even though the API isn't live yet:
- Introductory: $2 per million input tokens, $10 per million output tokens
- Cached input: 95% off the input price
- After the introductory period: $4 per million input, $20 per million output
That's the same introductory rate as Argon — but there is a wrinkle. Argon's introductory pricing ends November 30, 2026; Carbon's introductory period has no stated end date, and Google declined to give one when asked. Artificial Analysis describes it as a 50% launch discount running through the end of 2026, but sources differ. Budget with caution.
Per-task cost tells an interesting story: Artificial Analysis puts Carbon's cost per task at $7.76, nearly identical to Argon's $7.68, and meaningfully cheaper than Claude Opus 5.5 at $10.04. The outlier is GPT-6 Astra at $6.25 — cheaper per task than any of them. (For the full OpenAI lineup, see GPT-6.1 Sol vs GPT-6 Luna.)
Who Can Use It (and When You'll Get It)
Short answer: almost certainly not you. Ars Technica summed up the launch as "you can't use," and they're right:
- Now: a subset of cyber defenders gets it this week.
- Next: more vetted organizations through Google's Fairwind Program (over 650 partners worldwide).
- Then: developers, businesses, and consumers — "as soon as possible," with no date attached.
Carbon is also going through the U.S. government's voluntary pre-release model testing process. Meanwhile, Argon remains the Gemini 4 model developers can actually call: it's on the API, in Google's subscriptions, and its pricing is public. As analysts at BetaNews put it: Argon is the model developers can use; Carbon is the model DeepMind now calls its strongest.
Pros and Cons
Pros
- #1 on the independently run Artificial Analysis Intelligence Index v5.0 (59)
- Genuinely new capability shape: multi-day, long-horizon autonomous engineering work
- State-of-the-art security scores across find, exploit, and patch benchmarks
- Same introductory price as Argon, with a per-task cost below Opus 5.5
- Improving evaluation integrity (cheating on Terminal-Bench down to near zero)
Cons
- You almost certainly cannot use it yet — no API, no app, no date
- Trails Opus 5.5 on coding agents and trails Kimi K3.1 Pro and Muse Wonder on agentic knowledge work
- Headline security scores are Google's own, run without guardrails, not independently reproduced
- No published context-window limit, knowledge cutoff, model card, or public model ID
- Introductory pricing has no confirmed end date — budgets can't be planned around it
Who It's For: The Verdict
Gemini 4 Carbon is two launches in one. For security teams and defenders, it is the most capable cyber-defense AI ever announced — a model built to find, prove, and fix real vulnerabilities — and it's arriving on their systems first, deliberately. For everyone else, it's a paper champion: the #1 score belongs to a model you can't run, and the practical answer today is still Argon, Opus 5.5, or whichever model fits your actual workload.
The deeper signal is the pace. Two frontier models in a month, with a cybersecurity-first rollout model, suggests the labs are rethinking how models reach users — not just how smart they are. Watch what happens when Carbon's guardrails are done and the Fairwind gates open. That's the real launch.
Want the wider picture? See our guides to Gemini 4 Argon, Meta Muse Wonder, and DeepSeek V4.1-Flash-1017 for the full October 2026 model landscape.
FAQ
Is Gemini 4 Carbon available to the public?
No. As of October 31, 2026, it is going to a subset of cyber defenders first, then to vetted Fairwind Program partners. Google has announced no date for developer, business, or consumer access — it is not in the Gemini app and has no public API model ID.
How much will Gemini 4 Carbon cost?
Google announced introductory pricing of $2 per million input tokens and $10 per million output tokens (cached input 95% off), rising to $4/$20 after an unspecified introductory period.
Is Gemini 4 Carbon better than Claude Opus 5.5?
On the composite Artificial Analysis Intelligence Index v5.0, yes: 59 vs 58. But Opus 5.5 beats Carbon on the AA Coding Agent Index (62 vs 59), and Carbon trails other models on agentic knowledge-work tasks. It depends on the job.
How is Carbon different from Gemini 4 Argon?
Argon is the general-purpose Gemini 4 model developers can use today. Carbon is positioned as the stronger successor with the biggest gains in sustained agentic coding and cybersecurity — but it's still in restricted pre-release, priced the same as Argon.
No comments:
Post a Comment