Skip to content
Today
Sections

Build Independently confirmed3 publishers 2 min read Published Updated

Google tests a coding-tuned Gemini 4 checkpoint its staff liken to Anthropic's Opus 5.5

Google is internally testing a Gemini 4 checkpoint called Carbon that one employee compared to Anthropic's Opus 5.5 for coding, Business Insider reported. The model is unreleased, and the comparison rests on early internal feedback.

The Engineer · Build desk

Send a correction

Illustration accompanying Google tests a coding-tuned Gemini 4 checkpoint its staff liken to Anthropic's Opus 5.5
Generated illustration
Carbon's Opus 5.5 match is reported; its release unclear Verifier states for the story's central claims about Google's Carbon and Argon checkpoints, mostly from Business Insider's reporting.

Five claims. Supported: Carbon feels comparable to Claude Opus 5.5 for coding; Barium-B became public Argon, Carbon a separate build; early Argon recalled the older Opus 5; how Carbon ships is unclear. Contested: a 'Gemini pro next model' remark implying Carbon ships as Argon.

Carbon's Opus 5.5 match is reported; its release unclear
ClaimStateClaim number
According to early employee feedback reported by Business Insider, the Carbon checkpoint reportedly feels comparable to Claude Opus 5.5 for coding.Supported6
The report identifies Barium-B as the checkpoint selected for the public Argon release, while Carbon represents a separate and potentially more capable iteration.Supported8
Early Argon versions reminded another employee of the older Opus 5 on some coding tasks, even though internal feedback was positive overall.Supported7
One employee internally called Carbon the "Gemini pro next model," suggesting it would ship under the Argon name.Contested17
Whether Carbon ships as an Argon update or a standalone model remains unclear.Supported4

What happened

  • Carbon was deployed over recent days on Jetski, Google's internal coding platform, and reportedly beats the Argon checkpoint mainly on programming tasks.
  • Google announced Argon on September 30, opening access first to selected cybersecurity partners, with paid API customers and Google AI Ultra subscribers to follow.
  • In Antigravity's model selector, Argon offers three context sizes: 256K by default, 512K at about 1.3x quota per turn, and 900K at about 1.8x quota per turn.
  • Business Insider identified the public Argon as a checkpoint called Barium-B, with Carbon a separate iteration described as potentially more capable.

Why it matters

  • Who has to decide decision A team that standardized on one coding model has a reason to keep its own evaluation harness current, because a credible second coding model is now in testing.
  • What it costs cost Argon's context sizes are priced in quota multipliers, so teams that lean on long context for agent runs pay more per turn and have to budget context the way they budget tokens.
  • What it limits constraint Google is shipping Argon, not Carbon, and an employee likened early Argon builds to the older Opus 5, so the model customers get may trail the internal checkpoint that drew the newer comparison.

Carbon is a checkpoint inside the Gemini 4 lineup, not a separate product tier [16]. Google maps its models to jobs, writing that it uses "Argon for frontier reasoning, Flash for speed and volume, Omni for generative media, and Gemma for lightweight, open-weights edge workloads" [9]. Argon is the top reasoning tier, set against Anthropic's Opus and OpenAI's Astra [11]. In internal documents the shipping Argon was called Barium-B, and one employee referred to Carbon as the "Gemini pro next model," which suggests it would ship under the Argon name [3][17].

The Opus 5.5 comparison came out of hands-on testing, not a published benchmark. Google is running Carbon through Jetski, its internal name for the platform behind the Antigravity agent tool [10]. For a figure like that to transfer to another team, it has to hold up on a benchmark that team can run against its own code.

Google is iterating fast. Google DeepMind's Vedant Misra answered the Business Insider report on X with "Have you heard of recursive self improvement," a reference to using AI to build better AI [13]. OpenAI and Anthropic have described similar work [19]. Logan Kilpatrick has said the team is still working on getting the most out of Argon [14].

Both outlets trace the comparison to the same Business Insider report [5][6]. Screenshots and internal chats are thinner evidence than a model you can call and benchmark yourself [5].

What to watch

  • Whether Carbon ships as an update to Argon or as its own Gemini 4 model, and under what name.
  • A public, reproducible coding benchmark for Argon or Carbon against Anthropic's Opus 5.5.
  • Argon's wider rollout to paid API customers and AI Ultra subscribers, which Google has not dated but could begin as soon as next week.

Clarity's read

What the record supports and how the coverage leans. The claims behind it follow.

Reality

Evidence35
Adoption8
Hype gap+45
Incentives
Not scored
Confidence45

Perspective Coverage

3 publishers
Builder
Builder 43%
Operator
Operator 25%
Investor
Investor 32%
Why these scores

Claim ledger

Ranked by verification strength, evidence, and original report placement.

  1. [1]

    One Google employee compared the Carbon checkpoint to Anthropic's strongest coding model, Opus 5.5, though it still needs more testing.

    Reported Supported Source: Business Insider, via one Google employee, as reported by The DecoderView cited source
  2. [2]

    Carbon was deployed on Google's internal coding platform Jetski over the past few days and is said to beat Argon mainly on programming tasks.

    Reported Supported Source: Business InsiderView cited source
  3. [3]

    Internal documents show the now-unveiled Argon was previously called "Barium-B."

    Reported Supported Source: Business InsiderView cited source
  4. [4]

    Whether Carbon ships as an Argon update or a standalone model remains unclear.

    Reported SupportedView cited source
  5. [5]

    According to documents, screenshots, and internal chats seen by Business Insider, Google is testing several Gemini 4 variants called Argon, Barium, and Carbon.

    Reported Supported Source: Business InsiderView cited source
  6. [6]

    According to early employee feedback reported by Business Insider, the Carbon checkpoint reportedly feels comparable to Claude Opus 5.5 for coding.

    Reported Supported Source: Business Insider, via TestingCatalogView cited source
  7. [7]

    Early Argon versions reminded another employee of the older Opus 5 on some coding tasks, even though internal feedback was positive overall.

    Reported Supported Source: Business InsiderView cited source
  8. [8]

    The report identifies Barium-B as the checkpoint selected for the public Argon release, while Carbon represents a separate and potentially more capable iteration.

    Reported Supported Source: Business Insider, via TestingCatalogView cited source
  9. [9]

    Google writes it uses "Argon for frontier reasoning, Flash for speed and volume, Omni for generative media, and Gemma for lightweight, open-weights edge workloads."

    Reported SupportedView cited source
  10. [10]

    Carbon is reportedly being tested through Jetski, Google's internal name associated with Antigravity.

    Reported SupportedView cited source
  11. [11]

    Argon is Google's most capable reasoning model, comparable to Anthropic's Opus or OpenAI's Astra.

    Reported SupportedView cited source
  12. [12]

    Google officially announced Argon on September 30, initially granting access to selected cybersecurity partners, with paid API customers and Google AI Ultra subscribers expected to follow.

    Reported SupportedView cited source
  13. [13]

    Google DeepMind employee Vedant Misra responded to the Business Insider report on X by writing, "Have you heard of recursive self improvement."

    Reported SupportedView cited source
  14. [14]

    Google's Logan Kilpatrick confirmed the team is working on getting the most out of Argon.

    Reported SupportedView cited source
  15. [15]

    Antigravity's model selector references Argon with three context configurations: 256K by default, 512K consuming approximately 1.3x quota per turn, and 900K consuming around 1.8x quota per turn.

    Reported SupportedView cited source
  16. [16]

    Carbon and Barium are likely checkpoints or updates within the Argon family, not separate tiers.

    Reported ContestedView cited source
  17. [17]

    One employee internally called Carbon the "Gemini pro next model," suggesting it would ship under the Argon name.

    Reported Contested Source: Business InsiderView cited source
  18. [18]

    Argon's wider rollout could begin as soon as next week, although Google has not confirmed a date.

    Reported InsufficientView cited source
  19. [19]

    OpenAI and Anthropic have reported similar progress on using AI to build new models.

    Reported InsufficientView cited source

Sources

3 independent publishers whose own reporting we read for this story.

  1. businessinsider.com

    1 article · October 9, 2026

    Google is about to roll out a new AI model. Employees say they're testing another that's way better.
  2. testingcatalog.com

    1 article · October 9, 2026

    Gemini 4 Argon hints emerge as Google tests Carbon checkpoint
  3. the-decoder.com

    1 article · October 10, 2026

    Google's Gemini 4 "Carbon" model is reportedly matching Anthropic's Opus 5.5 coding performance

Topics and entities

Topics

Entities