Gemini 4 Carbon: Release Date, Price and What's New

Gemini 4 Carbon is out, but not for you — not yet. Google DeepMind introduced the model on October 28, 2026, calling it the company's strongest AI system to date, built around "state-of-the-art cybersecurity and long horizon capabilities." For now, access is limited to a small group of cyber defenders through something called the Fairwind Program, with wider developer and consumer availability promised only as "as soon as possible" — Google has not given a firm date. Pricing, however, is already set: $2 per million input tokens and $10 per million output tokens during the introductory period, climbing to $4 and $20 once that period ends, identical to the launch pricing of its predecessor, Gemini 4 Argon.

When is Gemini 4 Carbon's release date?

There isn't a single release date for Gemini 4 Carbon — there are two, and that's the source of most of the confusion already circulating. The model technically launched on October 28, 2026, but that launch is narrow by design. In its announcement post, Google says access starts "this week" with a subset of cyber defenders, expanding gradually to additional trusted partners through the Fairwind Program. Google is also participating in the U.S. government's voluntary pre-release model testing process, which is standard practice for a frontier model with security-sensitive capabilities.

For developers, businesses, and everyday consumers hoping to try Gemini 4 Carbon in AI Studio, the Gemini API, or the Gemini app, Google's own language is deliberately open-ended: broader availability is coming "as soon as possible," with no week, month, or quarter attached. Anyone searching for a specific public release date right now won't find one in the primary source — and treating a guess as a confirmed date would be inaccurate. The honest answer, as of this writing, is that the limited defender rollout is real and happening now, while the general release date is unannounced.

How much does Gemini 4 Carbon cost?

Pricing is one of the few numbers Google has nailed down, even while availability stays fuzzy. During the introductory period, Gemini 4 Carbon costs $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted 95% below the standard input rate. After the introductory period ends, the price rises to $4 per million input tokens and $20 per million output tokens — a doubling that mirrors exactly what Google did with Gemini 4 Argon, whose own standard-tier pricing jumps from $2/$10 to $4/$20 on December 1, 2026, according to the Gemini API pricing documentation. Google's Carbon announcement doesn't specify its own intro-to-standard cutoff date, so it isn't safe to assume it lines up exactly with Argon's November 30 cutoff.

In practice, that means Carbon is launching at price parity with Argon, not at a premium — despite Google describing it as the stronger model of the two. Whether that pricing holds once Carbon reaches general availability, or whether Google introduces a separate tier for the broader release the way it did for Argon's consumer-facing access, is one of the open questions in the next section.

What is Gemini 4 Carbon?

Gemini 4 Carbon is Google DeepMind's newest frontier model, positioned as the direct successor to Gemini 4 Argon, which Google announced just two weeks earlier, on October 13, 2026. Google frames Carbon less as a general-purpose upgrade and more as a specialist release: a model built to push two specific capabilities as far as possible — cybersecurity work and long-horizon agentic tasks — while carrying over the same multi-layered safety stack Argon introduced, combining model training safeguards, real-time misuse defenses, prompt-injection resilience, and chain-of-thought monitoring.

Koray Kavukcuoglu, SVP of Google DeepMind and the company's Chief AI Architect, wrote the announcement himself, which is itself a signal of how Google is framing Carbon: not as a routine point release, but as a model aimed at a narrower, higher-stakes audience first. For readers who followed Argon's launch — including its arrival as Google's newest flagship model in the Gemini app — Carbon is best understood as what comes after that consumer-facing release, built for defenders and long-running engineering work rather than for a mass rollout on day one.

State-of-the-art cybersecurity claims

Cybersecurity is the headline capability Google is pushing with Carbon, and the company backs it with several named benchmarks. On SEC-bench Pro, a vulnerability-finding benchmark, Google reports Carbon scoring 88.2%. On ExploitGym, which tests exploitation ability, it reports 54.8%. On CWE-bench, which measures patching quality against the Common Weakness Enumeration taxonomy, Google cites 71.0%. Google also says its internal testing shows more than double Argon's success rate on secure code generation, and that Google's own Chrome team found Carbon over 10 times better than another lab's large frontier model at producing high-quality security patches. Wiz, a cloud security firm, reportedly found Carbon leading on web penetration testing benchmarks as well.

Those are vendor-reported numbers from Google's own announcement and should be read as such — strong claims made by the model's maker, not independently audited scores. Readers who want background on how security researchers standardize and score vulnerabilities in the first place, independent of any one model's performance, can see our explainer on how vulnerability IDs and CVSS scores work, which covers the CVE system that benchmarks like CWE-bench build on conceptually.

Long-horizon agentic work

The second pillar of Carbon's pitch is sustained, long-running agentic work — the ability to stay on a task across many steps without losing the thread or giving up. Google says its internal teams have had Carbon run for days on multi-step engineering tasks, including root-causing and patching bugs across large codebases, persisting through dead ends rather than stalling out. That's a different kind of claim than a benchmark score: it's about endurance and recovery over long sessions, which is harder to quantify but is increasingly what separates "a model that can write code" from "a model you can assign a multi-day engineering ticket to."

Google also points to gains in web development and multimodal output quality. On WebDev Arena, a crowd-voted benchmark that ranks models on front-end code generation, Google says Carbon ranks #4 with an Elo score of 1768, up 89 points over its prior standing — a real improvement, if a more modest one than the cybersecurity numbers, and a reminder that Carbon isn't being pitched as a universal leader across every category.

Benchmarks: the Artificial Analysis Intelligence Index

Beyond the cybersecurity-specific tests, Google highlights Carbon's result on the Artificial Analysis Intelligence Index v5.0, a third-party aggregate benchmark that blends performance across reasoning, coding, and knowledge tasks into a single score. Google reports Carbon ranking #1 on that index with a score of 59. For context, Argon had claimed the top spot on a different aggregate, the Vals Index covering finance, coding, legal, and tax tasks, when it launched two weeks earlier — so Google now has two frontier models from the same family holding #1 rankings on two different third-party indexes within a single month, which is itself a notable (if slightly confusing) data point about how fast this product line is moving.

Readers who want the full three-way picture of where this generation of frontier models stands relative to competitors can see our earlier comparison of Gemini 4 Argon against GPT-6 Astra and Claude Opus 5.5; Google has not yet published an equivalent head-to-head for Carbon against those same rivals, so that comparison doesn't yet exist in primary-source form.

The Fairwind Program and why the rollout is narrow

The mechanism gating access to Carbon is called the Fairwind Program, described on Google DeepMind's own program page as the channel through which trusted partners — starting with cyber defenders — get early access to the model. Google frames this as deliberate: a model this capable at vulnerability discovery and exploitation is also, by definition, a model capable of helping the wrong actor do the same thing, so Google is giving the people whose job is to find and fix security holes a head start over a general release. The company says access will expand "gradually" to additional partners, without committing to a schedule.

Google also notes it is participating in the U.S. government's voluntary pre-release testing process for frontier models — a process several major AI labs have opted into for their most capable releases, typically involving government evaluators testing a model for serious risks (biosecurity, cybersecurity, and similar categories) before or shortly after public launch. Google doesn't attach a specific completion date to that process for Carbon either, which is part of why "later" is the most precise word currently available for the broader rollout.

How Carbon differs from Argon

It's worth being precise about what changed in two weeks, since Argon and Carbon are easy to conflate. Argon, announced October 13, 2026, was pitched as a broad, general-purpose flagship: strong across coding, finance, legal, and science tasks, with a 1-million-token unified input-output limit, a #1 result on the Vals Index, and same-day-ish rollout across the Gemini API, AI Studio, Antigravity for Ultra subscribers, and the Gemini app for Ultra and Pro subscribers. Carbon, by contrast, is narrower and deeper: it leads with cybersecurity and long-horizon agentic endurance specifically, reports safety and alignment improvements over Argon on several internal metrics, and — unlike Argon — is not rolling out broadly at launch.

On safety specifically, Google reports Carbon scoring 0.0% attack success on the Gray Swan Indirect Prompt Injection benchmark, down from Argon's already-low 0.7%. It also reports that high-severity misalignment alerts in live coding environments dropped by almost 80% relative to Argon, that unsupported claims in model outputs were cut roughly in half, and that "cheating" behavior on the Terminal-Bench 4.0 benchmark fell to 0.0% (with cheating attempts themselves down to 0.4%). Those are all Google's own internal measurements, published in the announcement and expanded on in DeepMind's separate Gemini 4 safety and security report.

Why the safety numbers matter for a security-focused model

The safety framing isn't incidental to Carbon's cybersecurity pitch — it's load-bearing. A model that's unusually good at finding and exploiting vulnerabilities is only safe to hand to defenders if it's also unusually resistant to being misused, manipulated via prompt injection, or pushed into unsupported or overreaching outputs. Google's reported drop in overreach (down 26%) and under-service (down 29%) relative to Argon reads as an attempt to answer exactly that concern: a cybersecurity model needs to be reliably calibrated, not just capable, since both overclaiming a vulnerability and missing one carry real costs for the defenders using it.

None of these figures are independently audited third-party numbers; they're Google's self-reported results, consistent with how Google benchmarked Argon's own safety profile two weeks earlier. That doesn't make them false, but it does mean they should be read as the maker's own claims about its own model, the same caveat that applies to the cybersecurity benchmark scores above.

What's still unknown

Three things aren't settled yet, and it's worth stating them plainly rather than guessing. First, there is no public release date for broader developer, business, or consumer access — Google's own wording is "as soon as possible," not a date. Second, Google hasn't said whether Carbon's eventual general-availability pricing will stay at $2/$10 rising to $4/$20, or whether a different tier structure will apply once it moves beyond the Fairwind Program, the way Argon's consumer access (Gemini app, AI Mode) came with its own subscription-tier gating separate from raw API pricing. Third, Google hasn't published a direct benchmark comparison between Carbon and current rival frontier models from other labs, the way it has for Argon — so any such comparison right now would be speculation rather than reporting.

What happens next

For the cyber defenders already inside the Fairwind Program, the immediate story is access: Google says it's "eager to see how the initial cohort of cyber defenders use this model" to find and fix real vulnerabilities, and the company's own benchmark claims will be tested against that real-world use well before most outsiders can verify anything hands-on. For everyone else, the next milestone to watch is simply whether Google attaches an actual date — a quarter, a month, anything more specific than "as soon as possible" — to broader API, AI Studio, or Gemini app availability. Until that happens, the two confirmed facts worth repeating are the ones Google has actually committed to: Carbon exists and is live for a limited group as of October 28, 2026, and it's priced at $2 per million input tokens and $10 per million output tokens during its introductory period, rising to $4 and $20 after that.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage