Major leak of Google's new Gemini 4 model, internal beta evaluation: Isn't this Opus 5.5?

Google is holding back a huge surprise!

While the full rollout of Gemini 4 Argon is still pending, the stronger Carbon version has already been secretly tested internally —

Even internal employees claim that its coding performance can already compete head-to-head with Opus 5.5 from its rival ~ (Is that really true?)

And it's not a small leak, it's a massive leak, because Carbon is just one of the models leaked this time.

Apart from the officially announced Argon, multiple R&D codenames including Barium and Carbon have emerged inside Google.

Argon, Barium, Carbon...

The full Gemini 4 lineup hasn't been released yet, but the periodic table of chemical elements has been almost exhausted. (doge)

What's more notable is that the trace of Gemini 4 Flash has also been spotted by sharp-eyed netizens!

Some community users shared the model identifier suspected to be in Google SDK: gemini-4-flash-preview, as shown below:

With the flagship model, the stronger internal test version, and the clues of the Flash series, this batch of information has pieced together a full set of the Gemini 4 family.

Google is releasing new models, iterating existing versions, and testing the next generation internally at the same time.

Who said the late bird can't get the worm!!

New Internal Version of Gemini 4 Exposed, Employees Say Its Coding Ability Can Compete with Opus 5.5

According to Business Insider, at least three internal codenames have appeared in the Gemini 4 series so far.

• Argon: The official Gemini 4 flagship model that Google has already announced.
• Barium: The previous series of versions that participated in internal tests, among which Barium-B was finally selected as the public version Argon.
• Carbon: The newly launched version for internal programming tests, which has received positive feedback from employees.

Let's first take a look at the newly leaked model — Carbon.

According to internal Google documents, employee chat records and related screenshots obtained by Business Insider, Google has been intensively testing new versions of Gemini 4 recently.

Among them, Carbon has been added to Google's internal programming platform Jetski in recent days for employees to experience.

As a result, before its official debut, it has already received a wave of "praise" from its own employees —

One employee who used it said: In programming tasks, Carbon feels almost as good as Claude Opus 5.5??

Another employee directly praised in the internal chat channel: Carbon is really good!

Some other employees mentioned that a new Gemini Pro model with the "Next" label performs quite well in use.

It is worth noting that the "Next" here is a test label used by Google for some new model updates.

From the relationship between the exposed models this time, there is also a small detail worth mentioning.

There is no strict one-to-one correspondence between Google's internal R&D codenames and the final product names released to the public.

For example, Barium-B finally became Argon.

This also shows that Google may evaluate multiple candidate versions at the same time during the internal testing phase, and then select the most suitable one for official release.

So it remains uncertain whether Carbon will eventually become an update to Argon, or be released as another independent model in the Gemini 4 series.

Speaking of which, the reason why Google employees compare Carbon with Opus 5.5 has its own background.

Because the information unearthed by BI this time also includes the internal evaluation of the early version of Argon.

According to reports, although the early version of Argon received good overall feedback from employees, some testers pointed out that its performance in some programming tasks still lags behind Anthropic's model.

Although Argon has achieved leading results in some software engineering benchmarks, it still falls behind Claude in tasks such as terminal operations —

For example, Argon scored 77.9% in DeepSWE v1.1, higher than the 74.2% of Opus 5.5 listed by the official, but in Terminal-Bench 4.0, Argon only got 57.4%, while Opus 5.5 scored 66.4%.

The different task types of the two tests also reflect the differences in the practical engineering capabilities of cutting-edge AI models.

Now that there is internal feedback saying Carbon's programming performance is close to Opus 5.5, it is natural to make people wonder if Google has carried out a new round of optimization for these shortcomings.

So we can look forward to Carbon, after all, they are all models from the same family, each generation is better than the last.

Flash Also Shows Up, Google Tightens Access Even for Free Users

Apart from Carbon, the news about Gemini 4 Flash has also become a hot topic recently.

In the past two days, discussions about Gemini 4 Flash have re-emerged on platforms such as X and Reddit.

Some developers shared screenshots of code suspected to be from Google SDK, where a very straightforward model name can be clearly seen: gemini-4-flash-preview

But judging from Google's update pace in the past few months, the Flash series has been playing an increasingly important role. Google just launched Gemini 3.8 Flash last month.

According to the official statement, while maintaining the positioning of low latency and low cost, this model focuses on enhancing the capabilities of long-cycle software engineering, multi-step reasoning and Agent workflow.

The flagship model is responsible for exploring the upper limit of performance, while the Flash series undertakes the larger-scale daily invocation demands, each serving its own purpose, which is a pretty good arrangement~

By the way, it is also worth mentioning that Google is recently adjusting the subscription tiers of Gemini products... (Here we go again!!)

According to The Verge, starting from October 9, free users' access to models such as Flash and Pro has been further restricted.

Previously, free users could switch between Flash Lite, Flash and Pro, but now the free tier only retains Flash Lite.

To continue using the standard Flash, users need at least a Google AI Plus subscription for $4.99 per month.

Paid AI Plus users are not spared either. Google says this tier will gradually lose access to the Pro model, and relevant users will be notified by email.

As for Pro and Deep Think, users need to subscribe to Google AI Pro for $19.99 per month, or the Ultra subscription for $99.99 per month...

To make things easier for everyone, here is a clear diagram to illustrate the new rules —

△

On one hand, Gemini 4 is constantly refreshing its performance records, on the other hand, the number of models available for free users is getting smaller and smaller.

It is literally that before the new models are fully served to the table, Google has re-priced the menu first...

Interested readers can stay tuned for the latest updates~

References:

[1]https://x.com/Haleeeemahh/status/2108666276218347970

[2]https://www.theverge.com/ai-artificial-intelligence/1005451/google-gemini-free-flash-lite-only

[3]https://www.businessinsider.com/google-employees-test-new-gemini-4-model-argon-barium-carbon-2026-10

This article is from the WeChat official account "QbitAI", written by Mengyao, and is republished with authorization from 36Kr.