Gemini 4 Carbon vs Argon: Same Price, Very Different Access

Google announced a new flagship model on October 28, 2026, just two weeks after the previous one reached most users. That short gap is the whole story of this comparison. Gemini 4 Carbon is, according to Google, the company's strongest model yet, with its sharpest focus on cybersecurity and long, multi-step work. Gemini 4 Argon, announced on September 30, is the model you can actually use today. Carbon is the model you probably cannot.

The two models share a family name, a starting price, and much of the same ambition. What separates them is not only performance but availability: Argon went through the same restricted-first rollout and came out the other side, while Carbon is still at the start of that path. The useful question is not which model wins a benchmark table. It is what changed, who can use each model right now, and which numbers deserve a caveat.

What happened

Google DeepMind announced Gemini 4 Carbon on October 28, 2026, in a post by Koray Kavukcuoglu, Google DeepMind's senior vice president and Google's chief AI architect. Google describes Carbon as its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities, and says it builds directly on Gemini 4 Argon, which Google says was launched two weeks earlier.

Argon's own timeline explains the comparison. Google announced Argon on September 30, 2026, initially for a small group of vetted cyber defenders through its Fairwind Program. Broader access followed on October 13, when Argon reached developers through the Gemini API and Google AI Studio, and reached paid consumer products. Carbon is starting that same journey now: Google says it will make Carbon available this week to a subset of cyber defenders, then gradually expand access to additional trusted partners through Fairwind. Developers, businesses, and consumers get it "as soon as possible," after Google gathers feedback and refines guardrails. No date has been given, and Google says it is taking part in the U.S. government's voluntary process for pre-release model testing.

Key features and changes

Carbon's headline claim is breadth: Google says it improved on Argon across all 17 rows of its published comparison table, with the most important gains in three areas: sustained engineering work, security testing, and agent reliability. All benchmark figures in this section are company-reported unless stated otherwise, and the security scores were, by Google's own note, evaluated without guardrails, so they should not be read as the performance a deployed customer will necessarily see.

Cybersecurity is where Carbon's lead is clearest. Google organizes the work into three stages: finding a vulnerability, checking whether it can actually be exploited, and producing a fix. On those tests, Google reports Carbon at 88.2% on SEC-bench Pro for finding real vulnerabilities (Argon: 83.4%), 54.8% on ExploitGym for validating exploitability (Argon: 42.0%), and 71.0% on CWE-bench for patching (Argon: 68.0%). The ExploitGym jump of nearly 13 points is the largest security gain. Google also reports, from internal benchmarks, that Carbon generates secure code successfully more than twice as often as Argon.

Long-horizon work is the second area of change. Google says its internal teams have seen Carbon run for days, persisting through dead ends on complex, multi-step engineering tasks such as tracing and patching subtle bugs across large codebases. In web development, Google reports Carbon gaining 89 Elo points to reach fourth place on WebDev Arena at 1768. On the Artificial Analysis Intelligence Index v5.0, a composite of independent evaluations, Google reports Carbon first at 59, ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57, and Argon at 50. Artificial Analysis has confirmed Carbon's first-place score on that revised index, while noting that the model has no publicly available API yet. Because version 5.0 uses a revised test suite, these scores should not be mixed with older index results.

Reliability is the third change, and it matters for anyone planning to hand a model a long task. Google reports that Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, reduced high-severity misalignment alerts in its internal coding environments by almost 80%, and roughly halved unsupported claims in internal testing. It also reports overreach down 26%, under-service down 29%, and a 0.0% attack success rate on Gray Swan's indirect prompt injection benchmark, against 0.7% for Argon. These are internal, company-reported results, not independent audits, but they point at the same goal: a model that stays inside its brief over a long run instead of guessing or fabricating its way forward.

The table below summarizes the head-to-head facts that are actually confirmed.

That last row is worth pausing on. Argon's documentation specifies a shared context of more than one million tokens, with extended generation beyond the default 65,536-token output limit. Carbon's announcement gives no equivalent API specification, so developers cannot assume those limits or tool features carry over unchanged.

Who it's for

Argon is for people who need a frontier model this week. Developers can call it through the API, prototype in Google AI Studio, and test it against their own workloads. Teams running long refactors or agent workflows can measure its real cost per completed task, rather than trusting a benchmark table. Paid subscribers can also meet it inside consumer products.

Carbon, for now, is for a much narrower audience: vetted cyber defenders working through the Fairwind Program, a controlled-access scheme that prioritizes governments, critical infrastructure operators, and core technology platforms. If you are a security team inside one of those organizations, Carbon is the model to watch, because its reported gains map directly onto finding, validating, and fixing flaws. If you are anyone else, Carbon is a preview, not a tool. Plan around Argon, and treat Carbon as a reason to re-run your evaluations when it arrives, not as a reason to pause work today.

Strengths and limits

Carbon's strengths, on Google's evidence, are real but uneven. The security gains are broad and consistent across all three stages of the defense lifecycle, which is rarer than a single standout score. The long-horizon improvements suggest a model built for work that does not fit in one prompt. And the reliability numbers, if they hold up outside Google, address the exact failure modes that make long agent runs expensive: shortcuts, fabrication, and drifting out of scope.

The limits are equally clear. First, access: a model with no public endpoint cannot be evaluated, budgeted, or depended on, whatever its scores. Second, evidence: the security benchmarks were run without guardrails, there is no technical report, and the narrowest margins in Google's table cannot settle which model suits a particular workload. Third, cost: identical unit prices do not mean identical bills. Carbon's announcement provides no data on token consumption per completed task, so no one can yet say whether it is cheaper per finished job than Argon. Argon's pricing, by contrast, is concrete and time-boxed: the introductory $2/$10 rate runs through November 30, 2026, then rises to $4/$20.

How to start

If you want to work with the Gemini 4 generation today, start with Argon, because it is the one you can open:

1. Open Google Gemini to see what the model does inside Google's consumer products, and check whether your subscription tier already includes it.
2. Prototype in Google AI Studio, where Argon is available for developers, and test it on a real task from your own codebase rather than a demo prompt.
3. For agent-style work, try Google Antigravity, the coding platform where Carbon is already in internal use and where Argon-powered workflows can be compared directly.
4. Before committing, measure cost per completed task at the introductory rate, and note the December 1 price rise to $4/$20 in your budget. Do not extrapolate Carbon's future cost from Argon's, since Carbon's introductory end date is unannounced.
5. If you are a defender in an organization eligible for Fairwind, that program, not the public API, is the route to Carbon today. Everyone else should watch for Google's wider rollout announcement and re-run benchmarks then, rather than planning around a date Google has not given.

Bottom line

Carbon versus Argon is not a normal upgrade story, because the newer model is not yet buyable. On paper, Carbon improves on Argon everywhere Google measured, most sharply in validating exploits and sustaining long engineering work, at the same introductory price. In practice, Argon remains the only Gemini 4 model you can evaluate, ship with, and budget for, and its introductory pricing window closes on November 30, 2026. Build and measure on Argon now, keep Carbon's claims marked as company-reported until independent testing catches up, and revisit the comparison the day Carbon gets a public endpoint. Until then, the most important difference between these two models is not a benchmark score. It is that one of them is here, and the other is still on its way.

Sources

Google Blog — blog․google VentureBeat — venturebeat․com Artificial Analysis — artificialanalysis․ai