A narrow lead on an independent index
Google announced Gemini 4 Carbon on 28 October, two weeks after Gemini 4 Argon, and called it its strongest model, with the biggest gains in cybersecurity, coding and long-horizon agentic work. The independent evaluator Artificial Analysis scored it 59 on its Intelligence Index v5, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57, and nine points above Argon’s 50.
That one-point lead is not spread evenly. Artificial Analysis gives Carbon the best score in four of the index’s nine evaluations — HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA — and places it second on its separate Coding Agent Index, where Carbon running in Google’s Antigravity CLI scores 59 against 62 for Claude Code with Claude Opus 5.5. On the same evaluator’s cost accounting, a Carbon run costs $7.76 per task against $10.04 for Opus 5.5 and $6.25 for GPT-6 Astra.
The clearest trade-off sits in AA-Omniscience v1.1, the index’s knowledge-and-honesty test. Carbon’s accuracy rose to 58% from Argon’s 48%, and its hallucination rate rose with it, from 4% to 11%. Artificial Analysis still records that as the second-lowest hallucination rate it measures, behind only Argon, and far below the 50% it reports for Claude Opus 5.5 and GPT-6 Astra at maximum reasoning.
The cybersecurity numbers are Google’s own
The scores Google leads with are self-reported. In its announcement, credited to Koray Kavukcuoglu, Google puts Carbon at 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploiting them and 71.0% on CWE-bench for patching root causes — against 83.4%, 42.0% and 68.0% for Argon. Google says the vulnerability work ran without guardrails, that internally Carbon has more than twice Argon’s secure-code success rate, and that Chrome engineers measured its patch quality at more than ten times that of an unnamed rival lab’s frontier model. None of those three claims has an independent counterpart yet.
Google also reports movement in its own safety evaluations: unsupported claims roughly halved against Argon, high-severity misalignment alerts in internal coding environments fell almost 80%, and on Gray Swan’s indirect prompt-injection benchmark the attack success rate is 0.0%, against 0.7% for Argon. On Terminal-Bench 4.0, Google says successful cheating fell to 0.0% and attempts to 0.4%.
Almost nobody can use it
Availability is the sharpest break from a normal frontier launch. Google says access starts this week with a subset of cyber defenders and widens to trusted partners through its Fairwind Program, with developers, businesses and consumers getting it later, “as soon as possible”, and no date attached. Artificial Analysis lists the model as having no public API. Argon, by comparison, has been generally available since 13 October.
Pricing is published anyway. Carbon costs $2 per million input tokens and $10 per million output during an introductory period, the rates Argon launched at, rising to $4 and $20 afterwards, with cached input discounted 95%. The context window is one million tokens.
What to watch
Google says it is taking part in the US government’s voluntary pre-release model testing process for Carbon. The next concrete dates are the end of the introductory pricing period — Argon’s runs to 30 November — and the point at which Fairwind access widens far enough that anyone outside Google’s vetted list can try to reproduce the cybersecurity numbers this launch is built on.
