One point, at a different setting

The headline number is 59, first place on version 5.0 of the Artificial Analysis Intelligence Index. The margin is one point over Claude Opus 5.5 at 58, with GPT-6 Astra at 57. Artificial Analysis measured Carbon at its high reasoning setting; the two rivals are listed at their maximum settings. That does not invalidate the result, and Carbon's nine-point jump over Argon's 50 on the same index version is a substantial move in two weeks. It does mean the ranking is a photo finish taken under uneven conditions, with no uncertainty estimates in Google's table to say how much of a one-point gap is signal. Readers comparing models should hold the version and the setting together: 59, on v5.0, at high reasoning.

The security numbers carry a condition

Google organised the launch around the full defense cycle, and the three headline scores map onto it cleanly: 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for checking whether a flaw can actually be exploited, and 71.0% on Collinear's CWE-bench for patching the root cause. The three evaluations cover different capabilities: identifying vulnerabilities, exploiting them in a test environment and producing patches. Together they offer a broader picture of Carbon's reported cybersecurity performance than any single benchmark, although the scores remain subject to the evaluations' stated conditions.

The condition is printed in the announcement itself, in parentheses: the external cybersecurity benchmarks were evaluated without guardrails. Google states that these external cybersecurity evaluations were conducted without guardrails. That is an important condition: the published scores do not establish how Carbon would perform with the restrictions, permissions and oversight applied in a deployed product. The figures describe performance under the stated evaluation conditions, not verified production outcomes. The internal evidence needs the same discount. Google says Carbon more than doubled Argon's success rate at generating secure code, and that its Chrome team saw it produce high-quality patches more than ten times better than another lab's large frontier model. The rival is not named and the underlying results are not published, so that comparison cannot be checked. It should be read as a claim, not a finding.

Reliability, measured two ways

Google's behaviour numbers point one way. In its internal testing, unsupported claims fell by roughly half against Argon, high-severity misalignment alerts in live coding fell by almost 80%, and on Gray Swan's indirect prompt-injection benchmark the attack success rate reached 0.0%, down from Argon's 0.7%. That last figure covers the attacks in that evaluated set. It is not a promise about every webpage, file or repository the model will ever read.

The evaluator's numbers point partly the other way. On AA-Omniscience, Artificial Analysis gives Carbon its highest score and a correct-answer rate of 58%, up from Argon's 48%. The same evaluation records a hallucination rate of 11%, up from Argon's 4%, where the rate counts incorrect answers against incorrect, partially correct and unanswered responses combined. These results come from different evaluations and measure different behaviors, so they are not a direct contradiction. Artificial Analysis reports that Carbon answered a larger share of questions correctly than Argon, while incorrect answers made up a larger share of its non-correct responses under the evaluator's hallucination metric. Google's separate internal findings of fewer unsupported claims do not resolve that distinction.

What is not in the announcement

There is no technical report yet. There is no parameter count, and no API specification setting out Carbon's context limit, inputs or endpoint, so Argon's documented specifications cannot simply be assumed to carry over. Google says its internal teams have run Carbon for days at a time on root-cause work in large codebases; that is a description of internal use, not a published limit a customer can design around.

Google's announced introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input discounted by 95%. VentureBeat and Unite.AI report that the announcement's pricing footnote sets later rates at $4 and $20, respectively. Those higher rates would double both headline token prices, but the announcement does not establish when they would take effect. Buyers should confirm the live pricing terms before budgeting for a production rollout. Argon's introductory period, by contrast, has a published end date of November 30, 2026.

Who can use it this week

Carbon is not generally available at announcement. Google says access begins with a subset of cybersecurity defenders, followed by gradual expansion to additional trusted partners through the Fairwind Program. Developers, businesses and consumers are promised access through Google products later, but no general-availability date has been announced.

Until that happens, Argon remains the Google flagship a customer can actually evaluate, with documented access and dated pricing.

Sources

  1. [1] Google, "Introducing Gemini 4 Carbon," blog.google, October 28, 2026 (primary source: Index v5.0 position, cybersecurity benchmark scores and the without-guardrails condition, 17-row comparison table, internal reliability figures, staged rollout and introductory pricing; the pricing footnote text was not included in the announcement page as read)Read source
  2. [2] VentureBeat, "Google unveils Gemini 4 Carbon, retaking the world's no. 1 spot on updated, harder benchmarks — but access is limited," October 28, 2026 (Artificial Analysis listings, Google comparison-table figures, AA-Omniscience detail, Argon timeline and pricing-deadline contrast; also reports, with Unite.AI, that the announcement's pricing footnote sets subsequent rates of $4 and $20)Read source
  3. [3] Artificial Analysis, "Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5," October 28, 2026 (evaluator's own results: Carbon 59 at high reasoning; Coding Agent Index 59 via Antigravity CLI against Claude Code with Opus 5.5 at 62; AA-Omniscience v1.1 score 54, accuracy 58% against Argon's 48%, hallucination rate 11% against 4%)Read source
  4. [4] Artificial Analysis Intelligence Index v5.0 (evaluator listings for Carbon, Opus 5.5, Astra and Argon, as reported above)Read source
  5. [5] Google DeepMind, Fairwind Program (rollout route named in the announcement)Read source