Gemini 4 Carbon: Google's New #1 AI Model at $2/$10

Google just released the model it calls its strongest ever, put it at the top of an independent AI leaderboard, priced it at $2 per million input tokens — and told nearly everyone they cannot use it yet. Gemini 4 Carbon, announced by Google DeepMind on October 28, 2026, is the company's new frontier model, and its rollout breaks the usual pattern: a subset of cyber defenders gets it first, vetted partners follow through the Fairwind Program, and developers, businesses and consumers have no date at all.

If you build on AI, three questions matter: what Carbon does, what it will cost when it reaches you, and whether it actually beats Claude Opus 5.5 or GPT-6 Astra — with Google's own numbers separated from the figures an independent evaluator has confirmed.

What Gemini 4 Carbon is

Google DeepMind describes Carbon as its "strongest model with state-of-the-art cybersecurity and long horizon capabilities," in the words of Koray Kavukcuoglu, the company's SVP of DeepMind and chief AI architect, who authored the announcement. The model builds directly on Gemini 4 Argon, which launched two weeks earlier. DeepMind's announcement says its researchers and models are now "working hand in hand to advance capabilities faster, while strengthening our safeguards in tandem."

Two things define Carbon. The first is cybersecurity: finding real vulnerabilities in software, validating them through exploitation, and turning them into working patches. The second is long-horizon work. Internally, Google says its teams watched Carbon run for days on multi-step engineering tasks, persisting through dead ends to root-cause and patch subtle bugs across large codebases and low-level systems.

The public basics: a 1-million-token context window, with text, image, video and speech input and text output (per Artificial Analysis).

The benchmarks — and which ones are Google's own

This is where careful reading matters. Carbon's loudest numbers come from Google's own announcement, and Google is upfront about one uncomfortable detail: the external security benchmarks were evaluated without guardrails. Treat everything below as the company's own account until independent testing repeats it.

Two items deserve a second look. First, the Intelligence Index placement is confirmed by the independent evaluator itself: Artificial Analysis published its own analysis on October 28, ranking Carbon first at 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. One caveat: Index v5.0 was only published on October 27 with a changed evaluation mix, so Carbon's 59 cannot be compared against older v4.3 scores as if they were the same scale.

Second, Google published a security comparison chart placing GPT-6 Astra at 85.4% on SEC-bench Pro and Argon at 83.4% (via WebProNews). Google also cites internal results: more than twice Argon's success rate at generating secure code, and a Chrome security team report that Carbon produced high-quality patches at more than ten times the rate of "another lab's large frontier model" — unnamed.

On alignment, the numbers are unusually specific: high-severity misalignment alerts in live internal coding environments dropped by almost 80%, overreach fell 26%, under-service fell 29%, and unsupported claims were cut roughly in half versus Argon. If those hold outside Google, they matter more than the leaderboard. A frontier model that fabricates half as often is a bigger deal than one Elo point.

The price: $2/$10 now, $4/$20 later — in USD, GBP and CAD

Google's announcement carries the pricing in one sentence plus a footnote. The introductory rate matches Argon's: $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off the input rate — that is $0.10 per million cached tokens. After the introductory period, the standard rate is $4 per million input and $20 per million output tokens, with cached input at $0.20 per million.

Read this part twice: Google has published no end date for Carbon's introductory period. Argon's intro pricing runs through November 30, 2026 before doubling on December 1 — I verified this on Google's own pricing page during research, and Carbon is not listed there at all. Artificial Analysis likewise notes the 50% launch discount with "no confirmed end date." Budget at $4/$20.

Conversions at November 1, 2026 market rates (£0.747 and C$1.397 per USD), verified at publication. Google bills in USD; your card rate will differ.

Two fine-print notes. First, the 95% cache discount applies at both tiers, so anything that reuses context — multi-turn agents, long coding sessions — runs at $0.20 per million input tokens or less at standard pricing. Second, the output price includes thinking tokens, and Carbon is a reasoning-heavy model. An independent analysis from Vals measured Carbon at $10.51 per test with a median latency of 47 minutes 45 seconds per task. This is a model that thinks for a long time, and you pay for every thought.

Cost per task: the number that actually matters

Token prices are the hourly rate; cost per task is the invoice. Artificial Analysis runs every model through the same set of Intelligence Index tasks and records what each run costs at standard list prices — folding in verbosity, thinking tokens and caching. Here is where Carbon lands at its top setting:

The reading: Carbon costs about 23% less per task than Opus 5.5 while scoring one point higher, but it is roughly 1.2 times the cost of GPT-6 Astra — our GPT-6.1 Terra review covers how Astra competes on price. Against Argon the comparison is stark: almost the same bill ($7.76 vs $7.68) for nine more index points. When Carbon reaches the public API, there is no price reason to keep building on Argon.

At the introductory 50% discount the arithmetic halves: roughly $3.88 per task (eesel.ai's calculation), undercutting Opus 5.5 and Astra at their max settings — but the discount has no published end date. And for small, high-volume workloads, GPT-6.1 Sol at ~$1.27 per task remains the value tier. Carbon is not a cheap model; it is a capable one priced below its closest peer.

Availability: defenders first, everyone else eventually

Carbon's rollout is the strangest part of this launch, and it is deliberate. Starting this week, Google is making the model available to "a subset of cyber defenders," then gradually phasing access to additional trusted partners through its Fairwind Program. The program's own page lists more than 650 partners, per WebProNews, with priority for governments, national cyber authorities, critical infrastructure operators and core technology platforms.

Google is also "actively engaged in the U.S. government's voluntary process for pre-release model testing" — meaning regulators are looking at Carbon before the public ever does. For developers, businesses and consumers, the company says access comes "as soon as possible," once it has gathered feedback and iterated on guardrails. No date, no waitlist, no public API model ID. It is not in the Gemini app on any plan, and as noted it is not on Google's pricing page yet.

There is a readable signal in the staging. Argon launched two weeks earlier and went straight to the paid API and consumer plans. Carbon — one generation stronger, with offensive-security skills as its headline — is gated behind vetted defenders and a voluntary government testing process. A model that scores 88.2% at finding real vulnerabilities and 54.8% at exploiting them is as useful to attackers as to defenders, and Google evaluated those benchmarks with guardrails off. The phased rollout is the safety strategy, not a marketing beat.

For security teams, early Fairwind access is a real edge: automated vulnerability discovery and root-cause patching is exactly the work Carbon's benchmarks describe, and agentic-security tools are becoming operational infrastructure, not research demos. For everyone else, the honest posture is patience: you cannot buy or test Carbon, so plan at $4/$20 and watch the guardrails once the model meets the open internet.

What to get right

Do not budget on the $2/$10 rate. The introductory period has no published end date, and Argon's precedent is a doubling. Build your cost model at $4/$20 — then the intro rate is a discount, not a surprise.

Model on cost per task, not cost per token. Carbon's $7.76 per Intelligence Index task at standard rates is the realistic number for heavyweight work, and the output side dominates because reasoning tokens are billed as output. If your workload is high-volume and light, a cheaper model like GPT-6.1 Sol (~$1.27/task) is the better fit.

If you are a security team, the Fairwind Program is the only door. There is no waitlist or public signup in the announcement, so start at the program's partner page rather than refreshing the Gemini API docs.

Do not treat the leaderboard as a purchase decision. The 59 on the Intelligence Index is independently confirmed by Artificial Analysis, but the index mostly tests text-based, English-language capabilities. It does not prove Carbon is the best model for your coding task, and Opus 5.5 still leads the Coding Agent Index.

Frequently asked questions

When will developers get access to Gemini 4 Carbon?

Google has not given a date. The announced order is: a subset of cyber defenders starting this week, then additional trusted partners through the Fairwind Program, then developers, businesses and consumers "as soon as possible" after feedback and guardrail iteration.

How much will Gemini 4 Carbon cost?

Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with cached input at 95% off ($0.10 per million). Standard pricing after the introductory period is $4/$20 per million input/output tokens. Google has published no end date for the introductory period.

Is Gemini 4 Carbon better than Claude Opus 5.5?

On the Artificial Analysis Intelligence Index v5.0, Carbon scores 59 versus Opus 5.5 at 58 — a one-point, independently measured lead. But Opus 5.5 still leads the Coding Agent Index, and Carbon costs $7.76 per index task versus Opus 5.5's $10.04. "Better" depends on your workload.

Have Carbon's security benchmarks been independently verified?

No. The 88.2% SEC-bench Pro, 54.8% ExploitGym and 71.0% CWE-bench figures are Google's own, evaluated without guardrails, and no outside evaluation has reproduced them at publication.

Is there a free tier for Gemini 4 Carbon?

Nothing published. Argon has no free tier on the paid API, and Carbon is not publicly available at all yet — so there is no free access to ask about. Watch Google's pricing page when the model goes public.

The bottom line

Gemini 4 Carbon is Google's strongest model on paper, with the independent leaderboard position to back the headline: 59 on the Intelligence Index, the strongest security-benchmark suite the company has published, and a per-task price that undercuts its closest peer by about 23%. But this is a controlled release, not a product you can buy — defenders first, partners next, everyone else on an undated "as soon as possible." The price to plan on is $4/$20, the benchmark to trust is the one Artificial Analysis ran, and the thing to watch is how quickly the guardrails hold up once the model meets the open internet.

Sources: Google DeepMind announcement (Oct 28, 2026); Artificial Analysis Intelligence Index v5.0 analysis (Oct 28, 2026); Google AI pricing page (accessed Nov 1, 2026); GHacks (Oct 31, 2026); WebProNews (Oct 28, 2026); AI Stack Crunch (Oct 28, 2026); Economic Times CISO (Oct 29, 2026); eesel.ai pricing analysis; agentriot; kingy.ai; thorstenmeyerai.com; top5apps.ai; oossa.