Google Rolls Out Gemini 4 Argon While Staff Test Carbon, a Checkpoint Some Say Rivals Opus 5.5

News Summary

Google is preparing to put its new Gemini 4 model, internally codenamed Argon, in front of more users, while employees are already trying a newer internal checkpoint called Carbon that some testers say is a clear step up, particularly for coding. The details come from a Business Insider report based on internal documents and employee chats, cross-checked here against secondary coverage of that report and of related reporting. Business Insider's original article could not be opened directly, so some details below come from aggregators and should be read as secondhand.

What Google has announced

Google announced Gemini 4 Argon on September 30, 2026. Published coverage gives the date but no clock time, so no timezone-specific timestamp can be stated. Access started with select cybersecurity specialists through a program called Fairwind. Paid API customers and Google AI Ultra subscribers are expected to be next, but Google has not published a date for broader availability.

Google also says thousands of its own employees already use the model internally for specialized coding, research and writing tasks.

What employees are testing

According to the Business Insider report as relayed by other outlets, employees have been trying a newer Gemini 4 version nicknamed Carbon. The report describes an internal naming trail that includes a model called Barium-B, which was reportedly chosen to become the model known publicly as Argon. Carbon appears to be a later checkpoint in the same Gemini 4 series rather than a replacement for Argon.

One tester said Carbon "feels like Opus 5.5" for coding, referring to Anthropic's Claude Opus 5.5, but added that more testing was needed. Some coverage frames Carbon as approaching Opus 5.5 through recursive self-improvement style training, though that framing comes from aggregators and has not been confirmed by Google.

How Argon has performed so far

Early independent testing placed Argon at 53 on the Artificial Analysis Intelligence Index, level with OpenAI's GPT-6 Astra. Claude Opus 5.5 still led that index at 58. Reports also say Argon hallucinates noticeably less than comparable models in those early tests.

A separate Bloomberg report, as relayed by secondary coverage, said some Google employees doubt Argon's real-world performance. They described strong scores on industry-standard benchmarks but weaker results on certain coding tasks. Two people said the model appears affected by "benchmaxxing", meaning optimizing for test scores rather than everyday usefulness. Google told Bloomberg it would be inaccurate to say the model underperforms in areas such as coding.

What is still unknown

Google has not confirmed that Carbon exists as a product, whether it will ship, or under what name. It is unclear whether Carbon would arrive as an update under the Argon name or as a distinct Gemini 4 family model. No independent benchmarks for Carbon have been published, so comments from internal testers remain anecdotal.

Why it matters

Frontier model releases now move through staged checkpoints, with internal codenames, limited early access and rapid iteration. The gap between benchmark scores and hands-on usefulness, especially for coding, is becoming the key way developers judge new models. Readers should treat internal tester impressions as early signals, not verified performance, until independent evaluations of Carbon appear.