Artificial Intelligence

Gemini 4 Carbon: The Model Google Is Testing Behind Argon

Google hasn’t finished rolling out Gemini 4 Argon, and employees are already testing something newer. According to an October 9 report, the newer model is Gemini 4 Carbon, an internal checkpoint one staffer says “feels like Opus 5.5” for coding. Google hasn’t confirmed Carbon exists.

That’s a lot of codenames for one week of news. Here’s what’s known, what isn’t, and why the coding angle matters.

Table of Contents

What is Gemini 4 Carbon?

Gemini 4 Carbon is an internal checkpoint of Google’s Gemini 4 model, tested by employees on Jetski, Google’s internal coding platform. A checkpoint is a saved version of a model at one stage of training or tuning. Google hasn’t announced Carbon or said when anyone outside the company might get it.

The report draws on internal documents and screenshots. Carbon showed up on Jetski in the days before the story ran, and the report says it’s possibly a new update to Argon. Google declined to comment, according to coverage of the report.

Gemini 4 Argon, Barium and Carbon: what each codename means

Codename	What it is	Status
Argon	The Gemini 4 flagship Google announced on September 30, 2026	Announced, limited access
Barium	A series of internal Gemini 4 test versions	Internal
Carbon	A newer Gemini 4 checkpoint on Jetski	Internal testing only

An internal document lists all three as Gemini 4 models under test. It also shows that the model Google launched as Argon carried the internal name Barium-B. Google’s internal names and public names don’t line up neatly, so “Carbon” says nothing about what the product will be called, or whether it becomes a product at all.

What Google employees say about Carbon

The employee who made the Opus 5.5 comparison also said they’d need to test Carbon more before trusting it. The report describes Opus 5.5 as Anthropic’s most powerful model for long-term agentic coding, so it’s a high bar to name.

The reaction to Argon was mixed. Early internal versions were generally well received, but one employee said they felt behind on some coding tasks and compared them to Claude Opus 5, an older Anthropic release.

My read: if both impressions hold up, Carbon would sit about one Anthropic version ahead of early Argon in coders’ eyes. That’s a big “if” resting on two anecdotes, and nobody has published a benchmark.

Why Gemini 4 Carbon matters for Google’s coding push

Argon landed on September 30 after months of delays. Reuters reported that Google dropped Gemini 3.5 Pro, which CEO Sundar Pichai had said would arrive in June. Google’s own release showed Argon ahead on some benchmarks and behind on two of the four coding tests it included. German outlet All-AI, citing Artificial Analysis, reports an index score of about 53 for Argon, level with OpenAI’s GPT-6 Astra.

The report says Anthropic and OpenAI have pulled ahead with developers doing complex engineering work. Put that next to Argon’s coding results and the Carbon chatter makes sense. Coding is where Google is weakest, so a checkpoint that coders like is the one Google would want to ship.

When will Gemini 4 Carbon be released?

There’s no release date. Even Argon hasn’t reached everyone. Google started with trusted cyber defenders through its Fairwind Program and says paid Gemini API customers and Google AI Ultra subscribers come next, with no timeline.

I’d expect Carbon, if it ships, to trail that. That’s inference, not reporting. The report also notes Google isn’t guaranteed to release the newer versions it’s testing.

What we don’t know about Gemini 4 Carbon yet

Whether Carbon ships as an Argon update, a separate Gemini 4 model, or not at all.
How it scores on any benchmark. None has been published.
Its pricing, context window and model ID.
Whether the Opus 5.5 comparison survives wider testing.