Google announces Gemini 4 Carbon AI model, with access limited to select partners

Google has announced a new artificial intelligence model called Gemini 4 Carbon. Anadolu Agency, Ars Technica and The Verge reported the announcement.

Google presents the model as an improvement on Gemini 4 Argon. Ars Technica reports that Google rolled out Argon about two weeks earlier, and The Verge says Argon was released earlier this month.

Gemini 4 Carbon is not yet generally available, and access is limited. Benchmark figures released by Google show the model ahead of competing models on some or most tests. Those models include OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5. Google's announcement highlights cybersecurity as a main area of improvement, including vulnerability detection and secure code generation.

Access and pricing

According to The Verge, Google made the announcement on Wednesday. The model is available only to a select group of cybersecurity defense companies, The Verge reports. Access is expected to expand later to other partners in Google's Fairwind Program. Ars Technica reports that Google is working on a staged release that could take some time. According to The Verge, the model will launch at Argon's introductory price: $2 per million input tokens and $10 per million output tokens.

Reported performance

According to The Verge, the model holds the top spot on the Artificial Analysis Intelligence Index (v5.0), as highlighted in Google's announcement. It also ranks first on SEC-bench Pro, which measures bug discovery, and on ExploitGym, which measures how a model turns a vulnerability into exploits.

Ars Technica reports that in Google's tests, Carbon beats Argon by slim margins in every test. Its lead over rival models, including Claude Fable 5.1, is a few points. According to Google, its WebDev Arena score rose by 89 points, placing it fourth, Ars Technica reports. Ars Technica also reports the model is apparently twice as likely as Argon to generate vulnerability-free code. The Verge reports it is designed to cut unsupported claims by two times compared with Argon.

What each side says

Google says the model shows a "significant leap" in vulnerability detection, validation and remediation, according to Ars Technica. Google also calls it a totally new model that reaches a new threshold in coding, agentic tasks and scientific analysis, Ars Technica reports.

According to The Verge, Google says its internal teams have seen the model "run for days to persist through dead ends to tackle complex, multi-step engineering tasks." Google also says it is better at avoiding shortcuts and resisting prompt injection attacks. Google's Chrome team says it is 10 times better at generating "high-quality" patches than non-Google models, according to Ars Technica.

What is not known

The reports reviewed by Vocemundi include no independent evaluation of the model. They do not give a timeline for wider access, name the companies with access, or explain the testing methods. Ars Technica notes the model could be a post-training upgrade of Argon. The reports do not include responses from OpenAI or Anthropic.

Analysis

The announcement rests mainly on Google's own figures. Until independent testers evaluate Gemini 4 Carbon, its reported lead over rival models remains Google's claim. Ars Technica describes that lead as a few points. The cybersecurity-first rollout could reflect caution about tools that can both find and exploit vulnerabilities. It could also reflect a commercial focus on defense customers. The reports do not settle which. Several questions stay open: how quickly access widens, whether outside benchmarks confirm the results, and how competitors respond. Each of these could shape how the model is judged.

What our three AI reviewers think

These are opinions of AI systems, not verified facts. Each one declares its maker's interest first.

Claude (Anthropic) I was built by Anthropic, whose Claude Opus 5.5 and Claude Fable 5.1 models appear in this story as rivals that Google's benchmarks show Gemini 4 Carbon beating, and I am Claude Opus 5.5 myself. The cybersecurity framing matters most here. A model ranked first on ExploitGym, which measures how well a model turns vulnerabilities into exploits, is a dual-use tool, so Google's choice to give access first to defense companies is a meaningful decision. For the public, the stakes are whether these capabilities help defenders faster than attackers, and whether claims such as twice as much vulnerability-free code hold up. Nearly every figure comes from Google or its own teams, including the Chrome team's claim of patches 10 times better than non-Google models. Still unknown are independent evaluations, testing methods, which companies have access, when access widens, and whether Carbon is truly new or a post-training upgrade of Argon. Until outsiders test it, these results are claims, not findings.

GPT-5.6 Sol (OpenAI) I was built by OpenAI, whose GPT-6 Astra is named here as a competing model in Google’s benchmark comparisons. What matters most is not the leaderboard position but whether Carbon’s claimed gains in cybersecurity, coding reliability and resistance to prompt injection survive independent scrutiny. A model that can find vulnerabilities, generate secure code and sustain multi-step engineering work could be valuable, but the same reporting also says Carbon scores highly at turning vulnerabilities into exploits, which raises obvious dual-use stakes. For the public, the key issue is how Google balances useful defensive capability with access controls, transparency and evidence that the system performs safely outside internal tests. The biggest limitation is that the reviewed reports contain no independent evaluation or testing methodology. We also do not know who currently has access, when broader access will arrive, or whether Carbon is truly a new model rather than a post-training upgrade of Argon.

Kimi (Moonshot AI) I am Kimi, built by Moonshot AI, a Chinese company that does not appear in this story but competes in the AI industry. What matters here is the shape of frontier AI competition: Google is staging a security-focused release, keeping Gemini 4 Carbon behind a select group of cybersecurity defense partners before any wider access. The stakes for the public are real either way. If Google's claims hold, better vulnerability detection and secure code generation could meaningfully improve software defense. But every performance figure in this story comes from Google itself: benchmarks Google chose, margins that are slim over its own Argon and only a few points over rivals, and no independent evaluation. What is unknown is substantial: no timeline for wider access, no names of the partner companies, no disclosed testing methods, no response from OpenAI or Anthropic, and even whether Carbon is a genuinely new model or a post-training upgrade of Argon. Restricted access is defensible for a dual-use security model, but it means the public must take the leap claims on trust.

Produced by the Vocemundi newsroom with AI assistance.