Google unveils Gemini 4 Carbon, with a benchmark lead and restricted rollout

What happened?

Google announced Gemini 4 Carbon on October 28, describing a frontier model aimed at software engineering, sustained agent workflows and cybersecurity defense. Access is restricted: Google says selected cyber defenders will receive it starting this week, followed by additional trusted partners through its Fairwind Program. Broader developer, business and consumer access is planned, without a firm date.

Google lists introductory pricing of $2 per million input tokens and $10 per million output tokens. Its stated standard rates after that period are $4 and $20 respectively. Artificial Analysis says the discount has no confirmed end date. This is an announced model with a phased rollout, not general availability today.

Source: Google ↗ · Source: Artificial Analysis ↗

The benchmark lead has boundaries

Artificial Analysis reports a score of 59 for Carbon with high reasoning on its Intelligence Index v5, placing it first in that composite evaluation. Its own analysis also makes the limits clear: Carbon is not leading every individual capability.

On the separate Coding Agent Index, Carbon using Antigravity CLI scores 59, behind Claude Opus 5.5 at maximum reasoning in Claude Code at 62. Those are evaluations of particular model-and-harness combinations, not proof that one model wins every coding task.

Source: Artificial Analysis ↗

Cybersecurity results need their test conditions

Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. Its announcement says the cyber benchmarks were evaluated without guardrails. The methodology document identifies SEC-bench Pro and ExploitGym as self-computed results, while CWE-bench results come from an external leaderboard evaluated by Collinear AI. These scores should not be mistaken for normal product settings or guarantees of real-world defense.

Source: Google ↗ · Source: Google DeepMind ↗

Why does it matter?

Our analysis: Carbon’s meaningful change is the combination of a named, official model and a strong independently reported composite score. That is a firmer basis for evaluation than the anonymous demonstrations that preceded the announcement.

The practical constraint is access. Most readers cannot yet test whether those gains translate into their own workflows. Until availability expands, the useful questions are who can obtain access, which evaluation setup matches the intended work, and what the eventual full-price economics look like.

Earlier coverage: October 26 reports, preserved below

The sections below retain our October 26 reporting and analysis of unverified Arena demonstrations. Google’s October 28 announcement now confirms the officially named model, but does not authenticate the earlier demo endpoints or their claimed comparisons. The original publication date and uncertainty around those earlier tests are preserved.

Source: WorldofAI ↗ · Source: Google ↗

October 26 archive: the reported Arena demonstrations

WorldofAI describes Arena coding demonstrations attributed to a Gemini 4 Carbon checkpoint, including an H200-themed 3D visualization and a voxel scene. The presenter praises the output but also identifies visible problems. LessCarbs has not independently identified the backend model or reproduced the tests.

For comparison, Google’s official Argon announcement identifies Gemini 4 Argon and describes a phased rollout. That release is a documented product announcement; it does not authenticate a different checkpoint merely because a demonstration is attributed to the same model family.

Source: WorldofAI ↗ · Source: Google ↗

October 26 archive: what the demonstrations could establish

Our analysis: an attractive generated scene can be a useful example of a workflow. It does not establish a broad coding lead. We would need the prompt, model identifier, reasoning settings, tool access, execution results and selection process to understand what the example measures.

Repeated trials matter as much as the best output. A comparison should show whether the code runs, whether it implements the requested behavior, how often it fails and how long the complete task takes. Without those details, it is easy to confuse visual polish with reliable engineering.

October 26 archive: why we withheld stronger claims

Our analysis: an experimental checkpoint may eventually become an important product. But a rumored model identity, a promising demo and a shipped release are separate claims, each needing its own evidence.

For now, the sensible takeaway is to watch for a named model, official access information and repeatable evaluation. The examples give developers something concrete to ask about. They do not establish a release date, confirm particular internal variants or prove parity with Claude or OpenAI’s models. Buyers should not plan a migration around that stronger interpretation.