Google reportedly tests Gemini 4 Carbon with stronger coding performance than Argon and comparisons to Claude Opus 5.5
Google is reportedly testing Carbon, a new internal version of Gemini 4 that employees have described as potentially more capable than the company's recently announced Argon model, particularly in complex software engineering tasks. The development emerged on October 9, 2026, less than two weeks after Google introduced Gemini 4 Argon as its latest frontier model for coding, professional reasoning, and cybersecurity.
According to internal documents and employee communications, Carbon has been available for testing through Jetski, Google's internal coding environment, with early users comparing its programming capabilities to Anthropic's Claude Opus 5.5. One employee described the experience as comparable to Opus 5.5, while acknowledging that further testing would be necessary before reaching a reliable conclusion.
The reported improvement is significant because Google has already announced ambitious performance results for Argon, including 77.9% on DeepSWE v1.1, a benchmark evaluating long-horizon software engineering. Nevertheless, the company is still preparing its broader commercial rollout, initially restricting access to selected cybersecurity partners while continuing to evaluate safety and reliability.
Carbon has not been officially announced, benchmarked publicly, or assigned a commercial release date. It remains unclear whether the internal model represents a new training checkpoint, a substantially revised Argon version, or a separate future release within the Gemini 4 family.
The distinction matters for developers evaluating Google's next generation of AI systems. Early employee feedback suggests that additional improvements are under development, but it does not establish how Carbon performs on standardized coding benchmarks, how much inference computation it requires, or whether its apparent advantages would survive deployment outside Google's internal infrastructure.
··········
GOOGLE IS TESTING CARBON WHILE PREPARING THE BROADER RELEASE OF GEMINI 4 ARGON.
Internal testing reportedly involves multiple Gemini 4 checkpoints, with Carbon emerging as a newer candidate after Google selected an earlier model for its initial Argon release.
Documents describing Google's development process identify several internal model names, including Argon, Barium, and Carbon. These labels do not necessarily correspond to publicly released products: an internal checkpoint called Barium-B was reportedly selected for the model Google announced publicly as Gemini 4 Argon, illustrating how the company distinguishes development versions from commercial model identities.
Carbon appears to belong to a subsequent stage of experimentation. Employees have reportedly been testing it through Jetski, an internal coding platform used to evaluate models on practical software development activities. The available information suggests that early testers have reacted positively, with some describing Carbon as a meaningful improvement over previous Gemini 4 versions.
Those observations are qualitative rather than experimental results. The number of employees who have tested Carbon, the tasks they performed, the amount of computation allocated to each session, and the criteria used to compare it with earlier models have not been disclosed.
An employee's assessment that a model feels comparable to Claude Opus 5.5 may reflect better code generation, more reliable tool use, stronger repository-level reasoning, or simply improved performance on the specific projects encountered during testing. Without detailed task records, these possibilities cannot be separated.
The reports also indicate that some earlier internal versions of Argon received mixed feedback on coding tasks. Although engineers generally viewed the model positively, at least one comparison placed its practical programming capabilities closer to an earlier Claude Opus generation.
Carbon could therefore represent an attempt to improve areas where developers found Argon less convincing. However, there is no evidence establishing that Carbon was specifically trained to correct those weaknesses, and Google has not disclosed whether the internal tests involve architecture changes, additional reinforcement learning, revised reasoning settings, or a new training run.
Google declined to comment on the reported tests, leaving the status of Carbon and its relationship with Argon unresolved.
........
Development detail | Current status |
Internal model name | Carbon |
Reported model family | Gemini 4 |
Testing environment | Google Jetski |
Reported specialization | Coding and software engineering |
Relationship with Argon | Unconfirmed |
Employee comparison | Claude Opus 5.5 |
Public coding benchmarks | Not available |
Architecture and parameter count | Not disclosed |
Context and output limits | Not disclosed |
API pricing | Not announced |
Public availability | Not available |
Official release date | Not announced |
........
One possibility is that Carbon will become an updated version of Argon following additional internal evaluations. Alternatively, Google could retain Argon as its initial commercial release and introduce Carbon later under a different name. A third possibility is that Carbon will remain an experimental checkpoint whose improvements are incorporated into another model without ever being released independently.
All three scenarios are technically plausible, but none has been confirmed.
··········
GEMINI 4 ARGON PROVIDES THE OFFICIAL BASELINE FOR EVALUATING CARBON'S REPORTED IMPROVEMENTS.
Google's September 30 announcement established measurable capabilities, commercial pricing, and an initial access strategy for Argon, none of which can yet be assumed to apply to Carbon.
Gemini 4 Argon is designed for sustained reasoning across complex professional tasks, with Google emphasizing software engineering, financial research, legal analysis, and autonomous cybersecurity operations.
The company has reported several concrete engineering applications inside its own infrastructure. Argon agents have assisted with large-scale migrations from C and C++ to Rust, including projects involving hundreds of thousands of lines of code, while other internal workflows have targeted data-center memory optimization and performance improvements in software components.
One documented example involves Google's libgav1 video decoder. Argon agents helped replace approximately 32,000 lines of SIMD-related code within an existing Rust port, using repeated profiling experiments and compiler analysis to produce a memory-safe implementation that Google reports is 2.7 times faster than the previous Rust version.
These examples illustrate the type of engineering work relevant to evaluating Carbon. A frontier coding model must handle more than isolated programming questions: complex tasks require understanding existing code, maintaining consistency across files, interpreting test failures, using development tools, and revising implementation strategies when initial approaches prove unsuccessful.
Google also reported that Argon achieved 77.9% on DeepSWE v1.1, although the result applies to the evaluation conditions disclosed for Argon and cannot be transferred to Carbon.
Another important architectural characteristic is Argon's unusually large 1-million-token output limit, increased from 64,000 tokens in earlier configurations. Google designed this capability to support extensive reasoning and execution trajectories without requiring the model to terminate or divide its output prematurely.
This is an output-token specification, not evidence that Carbon has an equivalent limit. The distinction is especially relevant when evaluating autonomous coding agents, because extended trajectories can consume substantial inference resources even if they ultimately produce relatively small code changes.
Argon's announced pricing also establishes a commercial reference point. Google plans to introduce the model at $2 per million input tokens and $10 per million output tokens, increasing to $4 and $20 respectively after the introductory period. Cached input tokens receive a 95% discount relative to the applicable input-token price.
The company has not indicated whether Carbon would inherit these prices, require additional reasoning tiers, or be offered as a separate premium model.
A technically stronger checkpoint can also be more expensive to serve if its improvements depend on additional reasoning tokens, more computationally intensive inference, or a larger underlying network. Consequently, developers cannot infer Carbon's eventual economic competitiveness from employee comments about its coding quality.
For now, Argon remains Google's documented Gemini 4 reference model, while Carbon is an unconfirmed development candidate.
··········
COMPARISONS WITH CLAUDE OPUS 5.5 REQUIRE MORE THAN POSITIVE EMPLOYEE FEEDBACK.
The reported comparison concerns the subjective experience of coding with Carbon, not a controlled evaluation demonstrating equivalent accuracy, reliability, or cost.
Claude Opus 5.5 provides a useful reference because Anthropic has positioned it for demanding coding and long-running agentic tasks. Such workloads place considerable pressure on a model's ability to maintain context, operate development tools correctly, and preserve consistency throughout multiple stages of an assignment.
However, saying that Carbon feels comparable to Opus 5.5 does not identify which of these capabilities was tested.
A model might perform particularly well at generating code for unfamiliar libraries but struggle when modifying a large existing repository. Another might produce fewer initial mistakes while requiring more tool calls to complete a task, increasing latency and inference expenditure despite an apparently smoother development experience.
An informative comparison must therefore separate at least four dimensions: correctness of the final solution, reliability during tool execution, resources consumed throughout the workflow, and the degree of human intervention required.
........
Evaluation dimension | Measurement required | Why it matters |
Coding correctness | Tests passed and regressions introduced | Determines whether generated changes actually work |
Repository-level engineering | Verified completion of multi-file tasks | Measures usefulness beyond isolated code generation |
Agent reliability | Tool failures, retries and recovery behavior | Captures problems encountered during autonomous execution |
Long-horizon reasoning | Successful completion of extended workflows | Tests consistency across multiple dependent steps |
Human intervention | Corrections and approvals required | Measures actual developer workload |
Inference efficiency | Tokens, latency and total execution cost | Establishes economic competitiveness |
Reproducibility | Repeated results across equivalent tasks | Distinguishes consistent capability from isolated successes |
........
For standardized evaluations, comparing Carbon with Argon and Claude Opus 5.5 would require the same benchmark version, compatible tool configurations, consistent retry policies, and transparent accounting of model-generated and tool-generated tokens.
Even under those conditions, published benchmark scores would provide only part of the evidence. An enterprise development team needs to know whether a model can work reliably with its own repositories, testing infrastructure, dependency constraints, and code-review requirements.
A useful practical experiment would involve assigning the same representative engineering tasks to each model, preserving equivalent access to documentation and development tools, and measuring the proportion of changes accepted after testing and review.
Cost per accepted software change could offer a more commercially meaningful measure than benchmark accuracy alone. It captures not only the price of inference but also unsuccessful attempts, repeated debugging cycles, and the additional engineering effort required to make a proposed solution production-ready.
Carbon currently lacks the public access and technical documentation necessary for such testing. Therefore, its apparent improvement over Argon and resemblance to Opus 5.5 should be interpreted as preliminary internal feedback, not as a validated change in the competitive ranking of frontier coding models.
··········
GOOGLE'S RAPID MODEL ITERATION COULD AFFECT ANTIGRAVITY AND THE ECONOMICS OF AI CODING AGENTS.
The commercial importance of Carbon depends on whether Google can convert internal model improvements into reliable capabilities available through its developer platforms.
Google's AI development strategy increasingly connects frontier models with environments in which they can perform sustained work rather than simply answer individual prompts. Coding assistants, automated engineering agents, and enterprise workflows require models capable of interacting with external tools, processing changing information, and completing tasks whose results can be verified.
Antigravity is particularly relevant because it provides an agentic software development environment in which models can operate on coding tasks. If Carbon becomes a commercially available Gemini model, such an environment would be a natural candidate for integration, although Google has not announced any Carbon-specific product plans.
The potential advantage is not limited to code generation. A more reliable coding model could reduce the number of failed implementation attempts, improve coordination between planning and execution, and lower the amount of human intervention required to complete repository-level work.
Those improvements could have substantial economic effects even without a dramatic increase in traditional benchmark scores.
Consider a hypothetical development agent that completes 70% of assigned tasks successfully at an average inference expense of $1 per attempt. If a replacement model increases successful completion to 85% while raising the cost per attempt to $1.15, the expected inference cost per successful task would decline from approximately $1.43 to $1.35, assuming comparable tasks, independent attempts, and no additional review costs.
The calculation illustrates why a more expensive model may still be economically preferable when its higher success rate reduces wasted execution. Conversely, a model that feels more responsive during interactive coding may offer limited financial benefit if it introduces additional regressions or requires more debugging.
Google has incentives to improve both performance and efficiency because agentic development environments can generate substantial inference demand. Unlike conventional chat interactions, agents may repeatedly inspect files, run tests, revise code, and evaluate intermediate results during a single assignment.
The competitive environment also extends beyond model quality. Google must coordinate model access, inference capacity, developer tools, reliability, pricing, and enterprise integration while competing with Anthropic, OpenAI, and other providers that are developing increasingly capable coding systems.
Carbon's reported existence suggests that Google is testing further improvements before Argon's broader commercial rollout has been completed. Whether that pace translates into a commercial advantage will depend on how quickly the company can validate and deploy new checkpoints without disrupting developers who need predictable model behavior.
Frequent model changes can create migration costs of their own. Organizations may need to revalidate prompts, agent workflows, evaluations, and security controls whenever a model's behavior changes materially, particularly when applications depend on complex sequences of tool calls.
For Google, a more capable successor would therefore need to deliver enough measurable improvement to justify its deployment and integration costs, rather than simply outperform an earlier checkpoint in internal demonstrations.
··········
CARBON'S RELEASE STRATEGY REMAINS UNCERTAIN AS GOOGLE CONTINUES ARGON'S CONTROLLED ROLLOUT.
The strongest evidence currently concerns the existence of internal testing, while Carbon's specifications, product identity, commercial terms, and eventual availability remain unresolved.
Google has already described a phased deployment strategy for Gemini 4 Argon, beginning with trusted cybersecurity defenders through its Fairwind Program. The company intends to gather additional evaluation feedback and strengthen safeguards before expanding availability to paid API customers, Google AI Ultra subscribers, and other users.
This controlled approach reflects Argon's advanced cybersecurity capabilities, including autonomous vulnerability discovery and remediation. Google has emphasized the need to evaluate potential misuse, prompt injection resistance, and the behavior of increasingly capable agents before broad deployment.
Carbon has not been included in this announced release schedule, and there is no basis for assuming that it has completed equivalent safety testing.
The possibility that Carbon represents a more capable model also raises additional validation requirements. Improvements in coding and tool use can increase the effectiveness of legitimate software engineering workflows, but models capable of conducting long-running autonomous operations must also be evaluated for unintended actions, security boundary violations, and failures to follow user instructions.
From a product-management perspective, Google has several possible choices.
The company could release Argon according to its current plans and introduce Carbon later as a separate upgrade. It could incorporate selected Carbon improvements into an updated Argon version, or it could retain Carbon internally while developing a different production checkpoint.
The final decision would depend on factors that are not publicly known, including performance consistency, evaluation results, serving costs, safety requirements, and the availability of sufficient inference capacity.
For developers considering the Gemini 4 ecosystem, the practical approach is to evaluate Argon using its published specifications and actual availability while treating Carbon as an unconfirmed future possibility. Delaying implementation solely because an internal checkpoint has received favorable feedback would introduce uncertainty without providing measurable technical or financial benefits.
The next developments worth examining are an official model announcement, a public or independently reproducible coding evaluation, confirmation of Carbon's relationship with Argon, and evidence of its availability through developer products or APIs.
Until those details emerge, claims of superior performance should remain qualified.
Google's reported testing nevertheless provides a useful indication of how quickly frontier model development is progressing. The company appears to be experimenting with additional Gemini 4 capabilities while its previously announced flagship is still completing a controlled release, suggesting that the distinction between successive model generations may become increasingly dependent on specific training checkpoints and deployment configurations.
Carbon's commercial significance will ultimately be determined by whether its reported coding improvements translate into more reliable completed engineering tasks, with reproducible results and competitive execution costs. Internal enthusiasm may indicate promising progress, but developers will need access to the model and comparable evidence before they can determine whether it represents a substantial advance over Argon or Claude Opus 5.5.
··········
FOLLOW US FOR MORE.
DATA STUDIOS
datastudios.org
