Image: Google

Google is already testing the next step for Gemini 4, and Argon isn't even out yet. Employees have been trying a new internal model called Carbon, according to Business Insider, which says it saw internal documents and screenshots. Google hasn't announced Carbon, and it declined to comment to the publication.

Google are already testing another Gemini 4 model called "Carbon" in Jetski

According to Business Insider, one employee says it "feels like Opus 5.5" for coding

Argon hasn't even dropped yet and they already have a better checkpoint, tf is going on rn pic.twitter.com/Omj8qIGv7g

— Lumina (@LuminaBench) October 9, 2026

Business Insider reports that Carbon recently appeared in Jetski, Google's internal coding platform. It may be a newer "checkpoint" of Argon, meaning an updated snapshot of the same model, but the publication couldn't determine whether Carbon will ship under the Argon name or as a separate Gemini 4 model.

The early reactions are what's drawing attention. One employee told Business Insider that Carbon "feels like Opus 5.5" for coding, a reference to Anthropic's top model for long agentic coding work, though they added that they'd need to test it more. "Carbon is really good!" another employee wrote in an internal channel.

That matters because coding has been Argon's sore spot. Business Insider says early Argon versions were generally well received inside Google, but one employee felt they were behind on some coding tasks, comparing them to the older Claude Opus 5. That echoes an earlier report that some Googlers felt Gemini 4 wasn't living up to its benchmarks in real work, a claim Google rejected.

Google's own numbers show where the gap is. In the chart it published at the Argon announcement, Argon beats Claude Opus 5.5 on DeepSWE v1.1, 77.9% to 74.2%, but trails it on Terminal-bench 4.0, 57.4% to 66.4%. So if Carbon really does feel like Opus 5.5, the clearest place to look for the improvement is that kind of hands-on, terminal-driven agent work. For now, that's one employee's impression, not a benchmark result.

The codenames also explain some of the confusion. Business Insider says an internal document lists a series of Gemini 4 models named Argon, Barium and Carbon, following the periodic table. Those names don't map neatly to what the public sees: the document reportedly notes that a model called "Barium-B" was picked to become the public Argon. In other words, the "Argon" that eventually reaches users may already be built on a later internal model, and Carbon could follow the same path.

There's no guarantee it ships, though. Business Insider stresses that Google isn't certain to release these updates. Argon itself is still limited to trusted testers and cyber defenders in Google's Fairwind Program, and Google has said the wider rollout will start with paid API customers and Google AI Ultra subscribers.

That rollout is already being prepared. As we reported earlier this week, Google's Antigravity coding app is getting ready for Gemini 4 Argon, with Ultra users first in line. If Carbon turns out to be a stronger Argon checkpoint, developers may end up getting a better model than the one Google showed off in September.

The bigger picture is pressure. Anthropic and OpenAI have pulled ahead with developers on complex engineering work, and Google is pushing on several fronts at once, including its new Gemini agent for work. Until Google says more, treat Carbon as a promising internal test rather than a product.