Google's Gemini 4 Carbon AI Targets Software Flaws, but Few Can Use It Yet

Google announced Gemini 4 Carbon on Wednesday, Oct. 28, calling it its strongest AI model yet. Most people and businesses cannot use it yet. Google said it will first go to a subset of cybersecurity defenders and reach developers, businesses and consumers later, with no date given.

The release comes two weeks after Google launched Gemini 4 Argon. Google's own blog post says Carbon builds on that model, and the company's performance claims have not been independently verified.

Who can use Carbon now, and who has to wait

According to Google, a subset of cyber defenders gets access starting this week. The company said it will gradually add other trusted partners through its Fairwind Program. It said it is still collecting feedback and refining safety guardrails before opening Carbon to developers, businesses and consumers "as soon as possible."

Google also said it is taking part in the U.S. government's voluntary process for testing AI models before release. Ars Technica noted that Argon, too, is on a staged rollout that could take a while.

What developers and companies could pay

Google said Carbon will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens. Cached input tokens are priced at 95% off the input price. Tokens are the small chunks of text that AI services use to measure usage and bill customers.

Anadolu Agency reported that the introductory pricing is temporary. It said standard pricing will later rise to $4 per million input tokens and $20 per million output tokens. Google's own post does not describe that later pricing.

What this means for ordinary users and businesses

For most consumers, nothing changes immediately, because the model is not in Google's public products. Businesses that build on Google's AI will have to wait for wider access, and Google has not said when that will come.

Google's emphasis on cybersecurity suggests where the early benefits may appear: tools that help security teams find, test and fix weaknesses in software. Google said Carbon is aimed at the whole defense process of detecting, validating and fixing vulnerabilities. Whether that translates into safer everyday software is not something the announcement shows.

What Google says Carbon can do

Google said Carbon tops version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. On external security benchmarks, evaluated without guardrails, it reported these results:

• SEC-bench Pro, finding vulnerabilities: 88.2%.
• ExploitGym, testing whether flaws can be exploited: 54.8%.
• CWE-bench, patching flaws: 71.0%.

Google said Carbon more than doubled Argon's success rate at generating secure code on internal benchmarks. It also said the Chrome team found Carbon performed more than 10X better than another lab's large frontier model at producing high-quality patches. On WebDev Arena, a web development leaderboard, Carbon gained 89 Elo points to reach fourth place at 1768.

The company described Carbon as built for long, multi-step engineering work. It said its own teams saw the model run for days on tasks such as tracking down and fixing subtle bugs in large codebases.

Reliability claims and how to read them

Google said Carbon cut unsupported claims by approximately 2x compared with Argon in internal testing. It said the model flags caveats or missing evidence rather than making things up. It also reported that overreach fell by 26% and under-service by 29% in multi-step tasks.

On Terminal-Bench 4.0, Google said successful cheating fell to 0.0%, with attempts at 0.4%. In live coding work with its own teams, high-severity misalignment alerts dropped by almost 80%, the company said.

Most of these figures come from Google's own testing. Ars Technica said the numbers show Carbon ahead of GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 in almost every test. It cautioned that the gap is only a few points, that every frontier model scores in the same neighborhood, and that benchmarks do not tell the whole story. Ars also said Carbon beats Argon by slim margins in every test Google cited, and that for all outsiders know it could be a post-training upgrade of Argon. Google presents it as a new model.

What happens next

Google said access will widen in phases, starting with cyber defenders and then more trusted partners. Broader availability will follow once guardrails are refined. The company did not give a date for a general release.

Around Prague, based on reporting by blog.google, Anadolu Ajansı, Ars Technica