Google Announced Gemini 4 Carbon. Argon Is Barely Out the Door.

Gemini 4 Argon started reaching Pro and Ultra subscribers on October 13. Fifteen days later, Google announced Gemini 4 Carbon.

That is a short honeymoon for a new flagship. The next model in the Google Gemini family promises harder engineering work, while ordinary subscribers are still getting started with the last one.

Google’s announcement puts a subset of cyber defenders first, then says it will phase in more trusted partners through its Fairwind program. Broader developer, business and consumer access remains undated.

For paying users, the immediate question is whether any of this changes the service they can use. For the wider industry, Carbon raises a more interesting question: how close are these systems to finishing difficult work with less supervision?

What Carbon is trying to do

Think of a bug that disappears when a developer tries to reproduce it. An assistant can explain the error message. An engineering agent has to investigate the project, follow evidence across files, try a repair and run checks. When the first approach fails, it needs to recover and keep the original problem in view.

Google says its teams have seen Carbon persist through multi-day engineering work, including subtle bugs in large codebases. Those internal observations describe the intended advance; public users still need a chance to test it.

For an everyday user, the appeal is follow-through. An assistant that can maintain a plan through setbacks could spare you the repeated explanations and corrections that make delegating a complicated task feel like managing an unusually forgetful colleague.

Argon already targets complex engineering and multi-step agents, according to its official documentation. Carbon builds on that direction. The question is how much further it can get before a person needs to intervene.

Why cyber defenders are first

Cybersecurity makes that question concrete. Finding suspicious code, proving that a flaw can be exploited and repairing it are separate jobs. A useful system needs to move between them without losing track of what the evidence establishes.

The SEC-bench Pro project evaluates vulnerability discovery and proof-of-concept generation. ExploitGym tests exploit development against real-world vulnerabilities. Google reports 88.2% and 54.8%, respectively. Its cyber evaluations ran without guardrails; neither benchmark’s own inspected page supplied a Carbon result for independent confirmation.

The distinction between those tests matters. A defender can use an authorized exploit demonstration to check whether a suspected weakness is serious enough to prioritize. A repair then has to close the weakness while keeping the software working. Google reports these as separate results, and they deserve to stay separate.

The same investigative ability could help someone attacking a system. Google’s Fairwind program aims to give vetted defenders a head start, prioritizing governments, critical infrastructure and core technology platforms. Permitted work is limited to defensive and research purposes, and partners cannot resell access.

Google says it is participating in voluntary US prerelease testing while refining safeguards. That is its stated deployment approach. The evaluations establish capabilities under particular test conditions, rather than a record of real-world attacks.

Independent tests back the coding and security gains

Artificial Analysis’s Intelligence Index v5.0 places Carbon first at 59 points with high reasoning. The next system, Opus 5.5 Max with default fallback, is on 58. AA’s launch analysis puts Argon nine points below Carbon. The index combines nine evaluations, so the lead describes that suite.

On AA’s separate Coding Agent Index v2.0, Antigravity CLI with Carbon at high reasoning scores 59, up from Argon’s 53 at the same effort level. Claude Code with Opus 5.5 Max and default fallback scores 62. The surrounding agent software matters: these are systems completing work, with different models, tools and effort settings.

There is also a useful operator check on defensive repair. Collinear’s CWE-bench v1 leaderboard lists Carbon with Antigravity at 71% programmatic pass@1, against Argon’s 68%. Each rollout has an hour to audit and repair a repository. A pass requires the exploit to stop working and the existing tests to keep passing.

That gives the security story substance. The benchmark checks whether a fix works under its rules, and the improvement over Argon is visible on the operator’s own leaderboard. Choosely has not tested Carbon.

Carbon and Argon have different strengths

Factual work complicates the upgrade story. AA’s launch analysis reports an 11% hallucination rate for Carbon at high reasoning, against 4% for Argon. The AA-Omniscience v1.1 definition uses incorrect answers as a share of non-correct responses, including partial answers and abstentions. It does not mean 11% of all Carbon answers are fabricated.

Carbon also answers more questions correctly and earns a higher overall knowledge score in that evaluation. Those results can coexist: a system can know more while making more bad guesses among the questions it fails to answer correctly. For fact-heavy work, both parts of the result deserve attention.

The practical difference today is access. Google’s app release notes document Argon for Pro and Ultra. Its API directory lists gemini-4-argon. Carbon remains absent from those inspected public listings, consistent with the announced restricted rollout.

The announced price comes before the public product

Google has announced these Carbon API rates in US dollars per million tokens. They are usage prices, separate from consumer subscriptions.

Cached input receives a 95% discount. Carbon’s introductory period has no announced end date. Argon’s API price schedule changes on December 1; that date belongs to Argon.

Matching token rates leave room for different job costs. AA’s own run supplies one early data point: at standard prices, Carbon cost $7.76 per Intelligence Index task and Argon $7.68, while Carbon scored higher. Those are weighted benchmark costs. The bill for your project will depend on its token use, retries and the human review it needs.

What to use while you wait

Keep using the access you already have. Google’s current plan notice names Argon for Pro and Ultra in the Gemini app, subject to eligibility and availability. A more expensive subscription carries no announced Carbon entitlement.

Developers can evaluate Argon through its documented paid API route. The pricing documentation excludes it from the API Free tier. Consumer subscription benefits and direct API billing are separate, so check the route you intend to use.

When Carbon opens more widely, use a few tasks from your actual work to judge whether it saves time. A coding agent should leave a working repair and tests you can inspect. A research assistant should leave claims you can verify. Those are more useful outcomes than a new name appearing in a model menu.

Choosely’s Take

Change nothing this week on Carbon’s account. Its first significance is for people building coding and security agents, where persistence and a working patch can be tested against clear outcomes. The independently published gains make Carbon worth evaluating when access opens.

For factual lookups, Argon’s lower hallucination metric is a reason to compare carefully before switching. Carbon’s higher knowledge accuracy keeps that from being a simple Argon-wins verdict. Choose it when it does your work better, under the access and safeguards you actually receive.