Gemini 4 Carbon vs Gemini 4 Argon: Release Date, Price and Benchmarks

Carbon is limited to cyber defenders; Argon is in the app and the API

Google's launch post says Carbon became available in its launch week to a subset of cyber defenders, with access widening gradually to additional trusted partners through Google's Fairwind Program. Google describes the order of the rollout itself: it is gathering feedback and iterating on guardrails, and taking part in the U.S. government's voluntary pre-release model testing, before making Carbon available to developers, businesses and consumers "as soon as possible". That is the only timing Google gives.

The table shows where each model stands on October 30, 2026. Every "not published" for Carbon means Google has said nothing yet, not that the model lacks the feature.

The Carbon post names no way to apply. Google's Fairwind page asks organizations to contact their Google Cloud account team and says Google prioritizes governments and partners most critical to societal resilience, but on October 30 it still describes access to Argon and never mentions Carbon. Outside that group, Google publishes no route to Carbon today.

Same price list, but only Argon's introductory rate has an end date

Read across the first two rows: Google's announced Carbon rates copy Argon's. The Carbon cached-input figure is ToolColumn arithmetic from Google's "95% off input" wording. The difference is timing. Argon's rates are dated on Google's pricing page, while Carbon's footnote says only that $4 and $20 apply after the introductory period expires, and Google says Carbon "will launch at" the introductory price, which suggests the period starts at a general launch that has not happened yet. Google also does not say which rate Carbon would carry if it launches after December 1, when Argon is already at $4 and $20.

For a budget, plan on $2 and $10 for early Carbon use and $4 and $20 after. An illustrative month of 50 million input and 5 million output tokens costs $150 at the introductory rate ($100 + $50) and $300 after it ($200 + $100) on either model, in ToolColumn arithmetic. Any real cost gap will come from how many tokens each model spends on a task, which Google has not published for Carbon.

Carbon leads Argon on all 17 rows of Google's launch table

Google's launch post and its Gemini product page carry the same 17-row table comparing Carbon with Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. Carbon leads Argon on every row and is first overall on 15. The rows below cover coding, computer use, knowledge work, long context and cyber work; the last column names the model Google's table puts first.

The gaps are uneven. Carbon's widest leads are on very long inputs (GraphWalks, 12 points), exploiting vulnerabilities (ExploitGym, 12.8 points), science tasks in a terminal (Terminal-Bench-Science 0.1, 64.8% vs 52.9%) and agentic coding and computer use (8.1 points on Terminal-Bench 4.0, 8.4 on Agents' Last Exam). On office automation (AutomationBench), chart reading (Chartography, 72.5% vs 71.6%) and the Vals Index v2.2 (70.4% vs 69.2%), Argon is within three points, so for document and analysis work Argon gives up little while Carbon stays closed.

Read these as Google's numbers. Google ran Carbon and Argon at their highest thinking settings, and computed several rows itself, including SWE-sweep, Agents' Last Exam, OSWorld, ExploitGym and SEC-bench Pro; its methodology credits the Intelligence Index and both Terminal-Bench rows to Artificial Analysis and the Vals Index to Vals AI. Unless the methodology says otherwise, results for the other labs' models are their self-reported figures; the Artificial Analysis and Vals rows come from those operators, and SWE-sweep competitors from its public leaderboard. Google says the cyber benchmarks were evaluated without guardrails, and the ExploitGym figures are its 10-hour runs (53.5% vs 39.5% with a 6-hour cap). This table uses different benchmark versions from Argon's model card, which lists Argon's Terminal-bench 4.0 result as 57.4%, so do not mix the two.

The launch post adds two Carbon-vs-Argon claims from Google's internal evaluations: more than twice the success rate at generating secure code, and about half as many unsupported claims. On Gray Swan's indirect prompt injection benchmark, Google reports an attack success rate of 0.0% for Carbon against 0.7% for Argon.

Build on gemini-4-argon now; Carbon has no model ID to test

Waiting for Carbon is no planned saving: its announced rates are Argon's, and Google has not said whether a launch after December 1 would still start at $2 and $10. Work built on gemini-4-argon today is what you will re-test when Google publishes a Carbon model ID, along with the context window and output limits it has not yet stated. In the Gemini app, Google has not said which plans will get Carbon, so moving from Google AI Pro to Ultra in the hope of Carbon access buys Ultra's higher Argon limits, not Carbon for now.