Gemini 4 Carbon reportedly matches Opus 5.5 at coding, even though Argon hasn't launched yet

Gemini 4 Argon hasn't reached mainstream users yet, but the conversation has already moved on to the next model in the family. According to Business Insider, as relayed by The Decoder, Google has a more powerful version codenamed Carbon. One company employee compared its programming abilities to Anthropic's Opus 5.5. There are no figures, no benchmark tables and no official comment from Google. What exists is a rumor from a single source and indirect signs of launch preparation.

What we know about Carbon

In essence, we have one sentence: the insider puts Carbon on par with Opus 5.5 specifically in coding. That qualifier matters. There is no mention of "general intelligence," multimodality or inference cost. The choice of comparison is not accidental: in recent years Anthropic has kept its reputation as the leader in software development and agentic workflows for programmers, so any claim of parity with its flagship draws attention.

Carbon is also positioned as a stronger variant within Gemini 4, which means Argon sits below it in this hierarchy. If the rumor is true, Google is preparing at least two tiers of the model, and the top one is aimed at the developer-tools market.

Signs of an imminent launch

The second part of the story is factual. New modes are appearing in the Gemini app and in AI Studio. This is usually what phased preparation looks like: Google first adds switches and operating modes to the interface, then opens access to the models themselves. The source doesn't specify which modes have been spotted or which model they are tied to, so a direct link to Carbon can't be drawn.

Argon itself also hasn't been rolled out widely yet. That means Google is going through the launch stage of the base model while already testing a heavier one. For a company of this size that is normal practice, but it also explains why leaks come from internal evaluations: employees are working with early checkpoints, which can differ significantly from the final release.

What is confirmed and what is not

• Confirmed by publications: Gemini 4 Argon is not yet widely available.
• Citing Business Insider: a version of Carbon exists that is stronger than Argon.
• Citing one employee: in coding, Carbon is comparable to Opus 5.5.
• Observed in products: new modes in the Gemini app and AI Studio.
• Unknown: release dates, pricing, context size, results on SWE-bench and other tests, API availability.

An assessment of "roughly on par with Opus 5.5" means little without a methodology. The employee may have compared the models on a handful of internal tasks rather than on an independent test suite. Moreover, in coding the gap between models is often determined not by a single benchmark number but by behavior in long agentic sessions: how well the model sticks to a plan, how it fixes its own mistakes, and how many tokens it spends on a task.

How teams should approach this

It is too early to move workflows to a model that isn't publicly available. A better approach is to lay the groundwork: assemble your own set of 20–30 real tasks from the team's repository, including refactoring, bug fixes and writing tests, and run it on your current model as a baseline. When Google opens Carbon or Argon in AI Studio, the comparison will take an hour, not a week.

It is also worth building the ability to switch model providers into your architecture in advance. If parity with Opus 5.5 is confirmed, price, limits and response speed will become the deciding arguments, and the teams that can compare options quickly will come out ahead. Until the official announcement, treat Carbon as a working hypothesis: a convincing one, but backed only by the words of one person.