What Is Gemini 4? Argon, Carbon and Every Variant Explained

Gemini 4 isn't one model — it's a family, and right now that family has two confirmed members: Gemini 4 Argon, the version almost anyone can use today, and Gemini 4 Carbon, a stronger but tightly restricted model built around cybersecurity and long-horizon engineering. Google has not shipped a standalone product called simply "Gemini 4," and it has not confirmed a Gemini 4 Pro, Flash, or "Hydrogen" variant. Everything you've read about those names is, as of this writing, rumor or third-party guesswork. Here's what Google has actually said, what it has quietly left unsaid, and what's still noise.

What is Gemini 4, exactly?

"Gemini 4" is Google DeepMind's name for its fourth major generation of Gemini models, not a single downloadable product. Google's own model-family page lists two Gemini 4 entries as of early November 2026: Gemini 4 Argon and Gemini 4 Carbon, plus a specialized offshoot called Gemini 4 Argon for Robotics. There is no separate page, model ID, or announcement for a plain "Gemini 4," and there's no confirmed Gemini 4 Pro or Gemini 4 Flash sitting alongside them the way Gemini 3.x had numbered Pro and Flash tiers. That's a real change in how Google is naming this generation, and it's a big part of why there's so much confusion online about which model is "the real" Gemini 4. Short answer: both Argon and Carbon are real, official Gemini 4 models. They just serve different purposes and have very different availability.

Gemini 4 Argon: the one you can actually use

Gemini 4 Argon is the model Google calls its flagship reasoning system for "complex software engineering, math and science problem solving, autonomous multi-step agents, and long decode." According to Google's own announcement post, Argon was written up by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google, and rolled out globally through October 2026 after an initial safety-testing window.

Confirmed access tiers, per Google:

• Consumers: Google AI Pro and AI Ultra subscribers get it in the Gemini app and AI Mode in Search; AI Ultra subscribers also get it in Google Antigravity.
• Developers: The Gemini API (paid tier only, model ID gemini-4-argon), available through Google AI Studio and Android Studio.
• Enterprises: The Gemini Platform, Gemini for Google Cloud, and Google Antigravity.
• Researchers: Google opened applications for 10,000 selected scientists to get six months of free access to Antigravity's highest usage tier.
• Cyber defenders: Early, guardrail-reduced access through the Fairwind Program, on a rolling basis.

On the technical side, the Gemini API documentation confirms Argon has a 1,048,576-token context window covering both input and output, a default 65,536-token output cap (extendable up to the full 1M for "long decode" tasks), and input support for text, image, video, audio and PDF — though output is text-only. Thinking is supported at low, medium, and high levels, and only implicit prompt caching is available; there's no explicit caching option. If you want the detail on how Argon rolled out inside the Gemini app specifically, Pandromeda covered Argon's release date, price and what's new in the Gemini app separately.

Google's self-reported benchmark highlights for Argon include a #1 ranking on the Vals Index (which spans finance, coding, legal and tax tasks), 87% on the Embodied Reasoning (ERQA) benchmark, and a Gray Swan indirect prompt injection attack success rate of just 0.7% across 15 adaptive attempts — numbers Google itself frames as its best to date, not independently audited results.

Gemini 4 Carbon: the cybersecurity specialist you (mostly) can't use yet

Two weeks after Argon shipped, Google introduced Gemini 4 Carbon, again authored by Koray Kavukcuoglu, describing it as "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." Carbon builds on Argon's safety and security stack and carries the same introductory pricing, but Google has been explicit that it is not yet broadly available: access began with "a subset of cyber defenders," with wider rollout to trusted Fairwind partners happening gradually. Developers, businesses, and ordinary consumers have no API, app, or product access to Carbon as of November 2026 — Google says it will extend access "as soon as possible" while it gathers feedback and tightens guardrails. Google also disclosed that it is participating in the U.S. government's voluntary pre-release model testing program with Carbon, which is a notable detail given the model's cyber-defense framing.

The benchmark numbers Google published for Carbon are considerably stronger than Argon's, particularly in security and coding:

• Artificial Analysis Intelligence Index v5.0: #1, score of 59.
• SEC-bench Pro (finding vulnerabilities): 88.2%, which Google calls state-of-the-art.
• ExploitGym (exploitation): 54.8%, also claimed as a new state-of-the-art.
• CWE-bench (patching): 71.0%.
• WebDev Arena: +89 Elo over Argon, reaching 1768 and ranking #4 overall.
• Terminal-Bench 4.0: cheating success rate down to 0.0%, cheating attempts down to 0.4%.
• Gray Swan indirect prompt injection: 0.0% attack success, versus Argon's 0.7%.

Google also reported internal and partner results it says show real-world impact: more than a 2x uplift in generating secure code over Argon on internal evals, roughly half the rate of unsupported claims, a 26% drop in "overreach" and 29% drop in "under-service," and an almost 80% reduction in high-severity misalignment alerts in live internal coding environments. The Chrome team reportedly saw more than 10x better patch quality compared with an unnamed rival lab's frontier model. All of these figures come directly from Google's announcement and have not been independently verified — treat them as Google's own claims, not third-party audits. For more on how the market and competitors reacted to Carbon specifically, see Pandromeda's rundown of what's confirmed about Carbon.

Argon for Robotics, and the names that aren't confirmed

Google's own Gemini model family page lists one more real Gemini 4 variant beyond Argon and Carbon: Gemini 4 Argon for Robotics, announced in October 2026 and aimed at physical-world, embodied reasoning tasks. Beyond that, the page does not list a Gemini 4 Pro or Gemini 4 Flash at all.

That hasn't stopped speculation. Several AI-model aggregator and tracker sites refer to a "gemini-4-flash-preview" model ID and attach it to an internal codename, "Hydrogen." We could not find this name, model ID, or anything resembling it on any Google-owned page, Google DeepMind documentation, or the Gemini API model list — according to the trackers reporting it, not according to Google. Treat "Hydrogen" and any Gemini 4 Pro/Flash release date as unconfirmed rumor until Google publishes its own model card or changelog entry for it.

Pricing and access, in plain terms

Both Argon and Carbon currently share the same published introductory pricing structure, per Google's Gemini API pricing page and both model announcements: $2 per million input tokens and $10 per million output tokens (including thinking tokens), with cached input priced 95% lower. That introductory window closes on November 30, 2026; from December 1, 2026 onward, standard pricing becomes $4 per million input tokens and $20 per million output tokens. Batch API pricing runs roughly half the synchronous rate. The catch for Carbon: none of this is actually usable by the public yet. The published Carbon price only applies once — and if — Google opens broader API or product access, which it has not committed to a date for.

For everyday users, the practical access question is simpler. If you have a Google AI Pro or AI Ultra subscription, you can use Gemini 4 Argon today in the Gemini app, AI Mode in Search, and (for AI Ultra) Antigravity. If you're a developer, you can call gemini-4-argon through the paid Gemini API tier right now. Gemini 4 Carbon, no matter what plan you're on, is not something you can sign up for — access is invitation-only through the Fairwind Program for vetted cybersecurity defenders and government testing partners.

How Gemini 4 differs from Gemini 3.x

Gemini 3.x was Google's prior generation, and it shipped with the more familiar Pro/Flash tiering: Gemini 3.1 Pro, plus a string of Flash releases (3.5 Flash, 3.5 Flash-Lite, 3.5 Flash Cyber, 3.6 Flash, 3.7 Flash, and 3.8 Flash/3.8 Flash Cyber) rolling out between roughly mid-2026 and September 2026, according to Google DeepMind's own model listing. Gemini 4 breaks from that pattern in a few concrete ways:

• Naming: Instead of numbered Pro/Flash tiers, Gemini 4 has shipped as named, purpose-specific models — Argon (general flagship), Carbon (cybersecurity/long-horizon specialist), and Argon for Robotics — at least so far. Google hasn't said whether a conventional "Gemini 4 Pro" or "Gemini 4 Flash" tier is coming.
• Security posture: Google explicitly measures and publishes resistance to indirect prompt injection attacks for Gemini 4 (0.7% for Argon, improved to 0.0% for Carbon on the Gray Swan benchmark) and frames Carbon's entire release around autonomous cyber-defense use cases — a sharper security focus than how prior Gemini generations were marketed.
• Staged rollout by capability, not just by price tier: Rather than releasing a full lineup at once, Google shipped Argon broadly first, then followed with the more powerful but higher-risk Carbon restricted to vetted partners and government testing. Google's own October 2026 AI recap frames Carbon as the next step after Argon rather than a competing release — an unusually cautious release pattern for Google's own flagship model.
• Benchmarks Google is willing to cite jumped sharply between the two Gemini 4 models themselves: Carbon's reported SEC-bench Pro (88.2%) and ExploitGym (54.8%) scores, and its #1 position on the Artificial Analysis Intelligence Index v5.0 (score 59), represent a larger self-reported jump over Argon than Google typically publishes between point releases within one generation.

One thing we could not confirm from any Google page: a direct, apples-to-apples context-window or pricing comparison between Gemini 3.1 Pro and Gemini 4 Argon. Google's DeepMind model page references 3.1 Pro only as a baseline in a partner (Box) evaluation, without publishing its own context-window figure alongside Gemini 4's.

What's confirmed vs. what's still rumor

Given how much noise surrounds "which Gemini 4 is the real one," here's the honest state of play as of November 8, 2026:

• Confirmed by Google: Gemini 4 Argon exists, is generally available, and has published pricing, a 1,048,576-token context window, and Google-reported benchmarks. Gemini 4 Carbon exists, was announced October 28, 2026, and is in restricted early access with published introductory pricing but no public release date. Gemini 4 Argon for Robotics exists and is a robotics-focused offshoot.
• Reported by Google but not independently audited: All of the specific benchmark percentages above (ERQA, SEC-bench Pro, ExploitGym, CWE-bench, Vals Index, Artificial Analysis Index, Gray Swan, Terminal-Bench) are Google's own self-reported figures from its announcement posts, not third-party lab results.
• Unconfirmed / rumor only: A Gemini 4 Pro or Gemini 4 Flash tier, a codename "Hydrogen," and any associated release date, pricing or context window for those names. These appear only on third-party AI-model tracker and aggregator sites, which we could not corroborate against any Google-owned source.

How Gemini 4 stacks up against rivals

Because Carbon is the stronger but far more restricted model, most real-world comparisons you'll see right now are actually comparing Gemini 4 Argon — the model people can actually run — against competing frontier systems, or speculating about Carbon using Google's own published benchmark numbers. Pandromeda has already broken down one of those head-to-heads in detail: see how Grok 4.7 compares with Gemini 4 Carbon on pricing, specs and access, since the Gemini 4 Argon access described above is gated by the same AI Pro/Ultra subscriptions that come up throughout that comparison.

Frequently asked questions

Is Gemini 4 Argon or Gemini 4 Carbon the "real" Gemini 4?

Both are real, officially named Gemini 4 models from Google DeepMind. Argon is the broadly available flagship; Carbon is a stronger, cybersecurity-focused model still in restricted early access. Neither is a fake or unofficial build of the other.

Can I use Gemini 4 Carbon right now?

Almost certainly not. Google has only opened access to a subset of vetted cyber defenders through its Fairwind Program and to a U.S. government testing partnership. There's no public API, app, or consumer path to Carbon as of November 2026.

How much does Gemini 4 Argon cost through the API?

Through November 30, 2026, it's $2 per million input tokens and $10 per million output tokens, with cached input priced about 95% lower. From December 1, 2026, standard pricing rises to $4 per million input tokens and $20 per million output tokens.

Is there a Gemini 4 Pro or Gemini 4 Flash?

Not officially, as of November 8, 2026. Google's own model-family page lists only Argon, Carbon, and Argon for Robotics under the Gemini 4 name. References to a Gemini 4 Pro, Flash, or a "Hydrogen" codename come solely from third-party tracker sites and are unconfirmed.

What's Gemini 4 Argon's context window?

1,048,576 tokens, shared across input and output, with a default 65,536-token output cap that can extend up to the full context for long-decode tasks, per Google's API documentation.

Why is Carbon's cybersecurity story such a big deal?

Google is positioning Carbon specifically as a tool for defenders — finding vulnerabilities, validating exploits, and generating patches — and is simultaneously testing it with the U.S. government. That combination of capability and caution is why Google is rolling it out far more slowly than Argon.

What's next for Gemini 4

The clearest open question is timing: Google has said it wants to bring Gemini 4 Carbon to developers, businesses, and consumers "as soon as possible," but hasn't committed to a date, and it's still mid-way through government pre-deployment testing. Watch for whether Google eventually fills out a more conventional lineup — a Pro or Flash-branded Gemini 4 model for cost-sensitive use cases — given that Gemini 3.x had several Flash tiers by the time Gemini 4 arrived. Until Google publishes something with its own name on it, any "Gemini 4 Pro," "Gemini 4 Flash," or "Hydrogen" claim should be read as a rumor, not a roadmap. We'll update this piece as Google's own pages change.

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