Gemini 4 Carbon Checkpoint Tested by Google Staff Ahead of Argon

What happened

Google has not finished releasing Gemini 4 Argon to the public, and its employees are already testing what may come after it. Business Insider reported on October 9 that staff have spent the past few days using a newer Gemini 4 build, codenamed Carbon, inside Jetski, Google's internal coding platform. The report cites internal documents and screenshots.1 The outlet described Carbon as possibly a new checkpoint of Argon, meaning a later snapshot of a model taken during training rather than a whole new system.1

The reaction inside Google was enthusiastic. One employee said Carbon "feels like Opus 5.5" for coding, comparing it to Anthropic's strongest model for long-running agentic coding, though they added that it needed more testing.1 Others wrote "Carbon is really good!" and said a "new Gemini pro next model feels kinda nice."1 Google declined to comment.28

The same internal document shows Google has tested a series of Gemini 4 models named Argon, Barium and Carbon. It also shows that a build called "Barium-B" was picked to ship publicly under the Argon name.1 It is still unclear whether Carbon will arrive as an Argon update or as a separate Gemini 4 model.1

The Flash question

This story has been widely framed as Google testing "the next Gemini Flash," and that label deserves some caution. None of the Carbon coverage says Carbon is a Flash-tier model. The only internal remark that hints at a tier calls it a "pro next" model.1 Still, the Flash framing makes sense given how Google has handled 2026. Flash has been the company's main vehicle for new models all year.

The pace has been fast. Gemini 3.5 Flash launched at I/O on May 19. Version 3.6 followed on July 21, then 3.7 on August 13 and 3.8 on September 2. Google DeepMind's Koray Kavukcuoglu said the company "took a little bit of a step back" from its Pro model to concentrate on Flash.2 Gemini 3.5 Pro was promised for June and never shipped. Axios reported that low morale at Google DeepMind was one reason for the delay.10

There is also some concrete evidence that a Gemini 4 Flash exists. In August, Google's tokenizer code was found to reference a gemini-4-flash-preview identifier, tied to the Gemma 4 tokenizer.29 An earlier rumor that a Gemini 4 Flash-Lite would appear in September turned out to be wrong.2 My reading is that a Gemini 4 Flash is very likely being built, but Carbon itself looks like a Pro-class checkpoint, and calling it Flash goes beyond the evidence.

Why Carbon matters: closing the coding gap

Carbon matters most because of where Argon reportedly fell short. Google announced Gemini 4 Argon on September 30 and said it led on several coding and knowledge-work benchmarks, ahead of OpenAI's GPT-6 Astra on some of them.10 On the same day, Bloomberg reported that some employees found the model weaker in real work than its benchmark scores suggested. Google called that report inaccurate.10 Business Insider's account fits with Bloomberg's: one employee said early Argon builds were behind on some coding tasks and compared them to Claude Opus 5, an older Anthropic model.1

Seen that way, the excitement around Carbon is a direct response to Argon's reported weakness. Going from "feels like Opus 5" to "feels like Opus 5.5" in a few weeks would be a real improvement in exactly the area where Google is under the most pressure: agentic coding. Several observers have pointed out that the two accounts can both be true, with Carbon possibly fixing the gap Bloomberg described.28

The speed also stands out. Argon was unveiled on September 30, and about nine days later staff had a reportedly stronger checkpoint. Google DeepMind employee Vedant Misra responded to the report on X by referring to recursive self-improvement, the idea of using current models to help build the next ones.30 Some outlets went further and said employees credited Carbon's gains to recursive self-improvement.2326 Those stronger claims go well beyond Business Insider's original report, which says nothing about how Carbon was built.1 Readers should treat the self-improvement angle as speculation, not a confirmed method.

How much a staff opinion is worth

The coverage agrees on the facts and disagrees on how much they mean. Some outlets treat Carbon as a turning point. BigGo, for example, argued that it shows Google's progress in agentic coding may be much faster than outsiders expected.26 More careful analysts point out that the whole story rests on a few chat messages: no benchmark, no model card, no context window, no price and no release date.27 One analysis called a positive internal opinion of a coding checkpoint "the weakest tier of evidence a builder can act on," because it comes from the people who built the model.30

The skeptics have the stronger case, and Google's own history in 2026 supports them. At I/O, Google said Gemini 3.5 Flash was a coding leap that beat Gemini 3.1 Pro on internal benchmarks.18 Google's own Android Bench results then placed it sixth, behind 3.1 Pro Preview, and the most expensive model on the list.18 By July, Google had replaced it with 3.6 Flash, and Ars Technica noted that 3.5 Flash had not lived up to Google's coding claims.19 Google said at the time that internal coding metrics showed a "massive, massive jump" from 3.1 Pro to 3.5 Flash.12 Strong internal numbers did not predict how developers would experience the model.

None of this means Carbon is overhyped. It does mean "feels like Opus 5.5" is a hypothesis that independent testing will have to confirm.

The release bottleneck

The bigger story may be how slowly Google gets these models to users. Argon first went only to Google's Fairwind Program, a group of vetted cybersecurity defenders, and Google said paying subscribers would be next after more testing.10 More than 650 vetted Google Cloud customers, government agencies and security partners have reportedly joined the program.23 As of October 11, one tracker said Google had published no API model ID, model card or knowledge cutoff for Argon.31

That may be changing. Google Cloud's model-lifecycle documentation now lists gemini-4-argon and a gemini-4-argon-cyber-permissive variant with an October 13 release date.32 Developers have also found Argon in the model picker of Google's Antigravity coding tool, with context options of 256K, 512K and 900K tokens.23 Taken together, these point to a broader rollout starting this week, at least on Google Cloud.

The gap between what Google tests internally and what users can access is unusually wide, and it is growing. Free Gemini app users were moved to Flash-Lite only as of October 9, and Flash now requires a paid AI Plus plan or higher.33 Gemini 3.8 Flash is still the newest Flash model available to developers, at an introductory $0.75 per million input tokens and $3.75 per million output tokens, a price that runs through the end of 2026.2 Meanwhile, the models Google staff are most excited about are two steps beyond what most customers can use.

Commentators have warned that if Google keeps its gated rollout while rivals ship faster, developers may lose patience. Some also worry that public versions could be scaled back compared with what employees test.26 Those warnings are opinion, but they reflect a real competitive problem. Axios described the past year as Google releasing cheaper Flash models while OpenAI and Anthropic pushed the frontier forward.10

The bottom line

Carbon shows that Gemini 4 is still changing quickly and that Google's teams think they have found a fix for Argon's reported coding weakness. It does not yet show that Google has caught Anthropic, and the evidence does not support calling Carbon the "next Flash." The most likely outcome is that Carbon, or something based on it, ships as an Argon update. A Gemini 4 Flash, already hinted at in Google's own code, would probably follow on the fast schedule the 3.x Flash models set.129

The next milestones will settle this better than any internal chat. They are whether Argon actually becomes broadly available through its listed API identifier, and whether a Carbon-derived model holds up on public leaderboards.2732 Until then, Carbon is a strong internal signal from a company that has already learned this year how much internal results can differ from real-world use.