Just now, Google's next-generation killer product Gemini 4 Carbon was exposed, and the major move of RSI is approaching Opus 5.5

Who could have imagined that before Google's Gemini 4 Argon is even released, leaks about its next-generation model have already emerged.

Just now, foreign media broke the news that Google's next-generation model under Gemini 4, codenamed Carbon, has been quietly launched for testing as a top-secret "killer weapon"!

It is reported that Google is currently testing the next checkpoint codenamed "Carbon" on its internal coding platform.

The performance of the new model has approached Opus 5.5, and multiple employees have revealed that Recursive Self-Improvement (RSI) is the key to achieving this progress.

Many internal beta employees exclaimed after trying it out: "It's so powerful!"

However, in the past two months, OpenAI has launched Astra, Anthropic has launched Opus 5.5, when will Google release a truly usable large model?

Argon hasn't been released yet, Carbon is tested in secret first! Google's internal team is working at a frantic pace

Today, exciting news comes: Google has recently added references to the Gemini 4 Argon model in Antigravity, covering three reasoning intensities: low, medium and high.

Now Carbon is also here, and it is said that there is also an internal coding tool called Jetsky.

To describe the current situation inside Google in one sentence: the progress is astonishing, and the team is pushing forward at an extremely fast speed.

On October 1, Google Gemini 4 Aargon was released, which is said to have achieved breakthrough improvements in benchmark performance in fields such as cybersecurity, law and finance.

But up to today, no ordinary user can actually use it.

Just when everyone was wondering if Google was doing another "PPT release", a leaked internal document and screenshots directly shocked all onlookers.

Is this news real? No one knows.

If Gemini 4 Argon, as shown in the picture, beats GPT-6 Astra and Opus 5.5 in multiple benchmark tests, Google will undoubtedly return to the top position.

However, under widespread public attention, the release of Argon has been delayed again and again.

It turns out that Google's engineers have no time to care about when Argon will be released to the public, because they are already completely immersed in testing a new version of the checkpoint called "Carbon"!

According to leaks, this mysterious Carbon model has been connected to Google's internal coding platform Jetski in the past few days.

This is no simple minor fix. An internal employee who just participated in the test excitedly told Business Insider: "On programming tasks, Carbon feels exactly like Opus 5.5!"

It should be noted that Opus 5.5 is currently the ceiling in the field of Agentic Coding.

Previously, many Google employees privately complained that the early version of Argon was slightly lagging behind in some complex coding tasks, and was only at the level of Claude Opus 5 at most.

Even Gemini 4 Argon has been reduced to a supporting test background.

In short, Gemini 4 never loses in benchmark scores, but its actual performance is far from satisfactory.

Influential figure @Token Gremlin used a vivid joke to mercilessly expose Google's real situation:

An ordinary day at Google:

Employee 1: "Boss, do we have a better model than Opus 5.5?"

Employee 2: "Based on my test results, no. Those fake benchmarks say we do, but we are far behind!"

Sundar Pichai: "Why don't we just tell everyone it's better than Opus 5.5, but don't let anyone test it?"

But the emergence of Carbon has directly closed this gap!

The chat records of employees on the internal Slack show this style ——

"This thing feels completely comparable to Opus 5.5!"

"Carbon is surprisingly powerful!"

"The new Gemini pro next model works perfectly!"

In short, Google is iterating the Gemini 4 series at a frantic speed.

There is even an internal document showing that in addition to Argon and Carbon, they are also testing a model called "Barium".

It seems that Google is obsessed with naming models after elements in the periodic table.

Core big move exposed! How powerful is RSI?

How could Google make such a huge leap in model performance in such a short period of time?

The answer is RSI.

Multiple Google employees mentioned this term in their leaks, and regard it as the key to Carbon's major progress at present.

It allows AI to discover errors in the code by itself while completing tasks, generate better solutions on its own, and then internalize these smarter logics into its own capabilities.

With each iteration, it becomes smarter than its previous version; and the smarter it gets, the faster it evolves!

When an AI with Opus 5.5 level programming capability optimizes itself 24 hours a day, it will usher in an exponential explosion of capability.

However, it seems that Google is not the only company that has implemented RSI.

Industry insiders pointed out in analysis that Google has apparently streamlined its internal organizational restructuring recently, and begun to focus on solving key technical pain points.

A well-known observer wrote on X: "I hope to see Google release a checkpoint that is significantly stronger than competitors every month, so that I can agree that the invincible Google has returned."

And the amazing performance of Carbon in internal testing now seems to be proving: RSI has been launched inside Google.

Is it a "real champion" or a "mediocre work"?

However, while Google's internal team is extremely excited, the outside world holds completely different attitudes.

After all, Google has a lot of bad records of "weakening the model before release".

Discussions about the exposure of Carbon on X have been polarized.

Optimists like netizen @Haleeeemahh directly posted the benchmark test screenshots from the previous conference, supporting Google: "Gemini 4 Argon has actually beaten Opus 5.5! It seems that the Google team really didn't exaggerate at the last event. Now there is Carbon which is comparable to Opus 5.5, do you still think Google will always be the second best?"

But more people have started mercilessly mocking.

Tech expert @HistoryGPT raised a sharp question: "Weeks after your competitor released their top model, you came up with an internal beta version that is just as good as theirs... Does this really prove that you have mastered the so-called RSI? I'm not sure."