Gemini 4 Carbon Reportedly Rivals Opus 5.5 at Coding

Gemini 4 Carbon Reportedly Rivals Opus 5.5 at Coding

Google has not released Gemini 4 "Argon" yet, and reports about its successor are already circulating. A new variant called Carbon is being tested inside the company, and at least one Google employee thinks it can compete with Anthropic's best coding model.

The details come from Business Insider, which says it reviewed documents, screenshots and internal chats. According to that reporting, Google is testing several Gemini 4 variants under the code names Argon, Barium and Carbon. The evidence is unofficial and early, but it gives an unusual view of how quickly Google is iterating on its next frontier model.

Carbon lands on Google's internal coding platform

Over the past few days, Carbon was deployed on Jetski, Google's internal coding platform. Business Insider reports that it outperforms Argon, mostly on programming tasks.

One employee compared Carbon to Opus 5.5, Anthropic's strongest model for coding. That comparison comes with a caveat: the model still needs more testing before anyone can draw firm conclusions.

The comparison stands out when set against earlier feedback on Argon. On some coding tasks, early Argon versions reminded another employee of Opus 5, Anthropic's older model. Internal reactions to Argon were still positive overall. Still, the gap between "feels like Opus 5" and "feels like Opus 5.5" suggests Carbon is a meaningful step up for software work, at least according to the people using it inside Google.

What Argon actually is

There was some confusion about Argon's role. Early rumors described it as a Flash model, a version tuned for speed rather than top performance. Google cleared that up when it announced its Gemini agent for Google Workspace.

In that announcement, Google assigned each model family a specific job. In the company's words: "Argon for frontier reasoning, Flash for speed and volume, Omni for generative media, and Gemma for lightweight, open-weights edge workloads."

That puts Argon at the top of Google's lineup. It is Google's most capable reasoning model and sits in the same class as Anthropic's Opus and OpenAI's Astra.

Carbon and Barium are therefore most likely checkpoints or updates within the Argon family, not new tiers of their own. A few details support this:

  1. One employee referred to Carbon internally as "Gemini pro next model," which points to a release under the Argon name.
  2. Internal documents show that the Argon model Google has now unveiled previously carried the name "Barium-B."
  3. The code names appear to track successive versions of the same model line rather than separate products.

It is still unclear whether Carbon will ship as an Argon update or as a standalone model.

A faster update cycle

Google has already moved quickly with its Flash models, and the Gemini 4 code names suggest the same fast cadence. This could mean Google has become better at using AI systems to help build their successors.

Google DeepMind employee Vedant Misra seemed to hint at this. Responding to the Business Insider report on X, he wrote: "Have you heard of recursive self improvement." The phrase describes AI being used to create better AI. OpenAI and Anthropic have both reported similar progress in their own work.

Google's apps are getting ready

Google has not announced an official launch date for Gemini 4, but its products are already changing in ways that look like preparation.

  • The Gemini app now offers an "Automatic" mode for the 3-series models.
  • Users can set reasoning intensity anywhere from low to high, a change that comes as Google has also restructured its Gemini tiers.
  • Google AI Studio, the company's developer tool, has added a new "Ultra" mode that promises "advanced skills and tools."

Google's Logan Kilpatrick confirmed that the team is working on getting the most out of Argon.

Our Take

Leaked internal impressions are not benchmarks. A single employee comparing Carbon to Opus 5.5 is only a data point, and the report itself says more testing is needed. Readers should treat the claim as a signal of direction, not a verdict.

The direction still matters. Coding has become a key battleground for frontier labs, and Anthropic's Opus models have served as a reference point there. If Google can close that gap within the Argon family, developers choosing a model for coding agents may have a stronger alternative from Google than they had before.

The more interesting thread may be the speed. Misra's comment, together with similar statements from OpenAI and Anthropic, fits a wider industry story about AI speeding up model development. Google's own research on self-improving agents points the same way. The next things to watch are whether Carbon ships under the Argon name, how it performs in independent testing, and whether Google's short release cycles hold up once Gemini 4 is public.