Google is internally testing the new Carbon model to enhance its programming capabilities and compete with advanced Claude models, amid widespread anticipation for the rollout of the Argon variant to users.
- Getsky leaks and comparisons with Anthropic models
- Multiple internal codenames and development pathways
- Details of the Argon model and the Fairwind cybersecurity program
- Internal debate at Google regarding model efficiency and technical leadership
- Frequently asked questions
Getsky leaks and comparisons with Anthropic models
Google engineers and employees are undergoing intensive trials and tests to evaluate an advanced internal version of its fourth-generation artificial intelligence models, bearing the codename “Carbon.” Initial evaluations showed that the new model’s capabilities in coding and computer programming now match the performance of the strongest competing advanced models from Anthropic. This news comes at a time when the flagship publicly announced model, “Gemini 4 Argon,” remains restricted and unavailable to most users more than a week after its official unveiling.
The news outlet Business Insider was the first to reveal these secret tests, based on leaked internal documents and screenshots from the company’s private software environments. Sources explained that the “Carbon” model was recently added to Getsky, the internal platform dedicated to developing and writing code for Google engineers. The report quoted an employee as saying that the programming experience using “Carbon” is very similar to working with Anthropic’s “Opus 5.5” model, which is considered a superior benchmark in long-term autonomous programming tasks, while emphasizing at the same time that the model still needs additional testing and improvements. In internal employee chat channels, engineers wrote words of praise such as: “Carbon is truly amazing” and “The upcoming new model delivers smooth and excellent performance,” while the company declined to officially comment on these leaks.
These comparisons reflect an overwhelming desire among Google’s technical team to regain the initiative in the programming and complex application development sector, after competitors’ tools gained widespread popularity among professional developers over the past months.
Multiple internal codenames and development pathways
One internal document revealed that research teams at Google tested a sequential series of fourth-generation models bearing chemical element names, including: “Argon,” “Barium,” and “Carbon.” It is customary in major tech companies for internal codenames not to match the commercial names under which products are later launched in global markets.
Indeed, the model chosen for public release under the name “Argon” was known inside the company’s labs as “Barium-B.” Until this moment, it remains unclear whether the “Carbon” model will reach the market as a developmental update under the name “Argon,” whether it will be launched as a standalone model within the Gemini 4 family, or whether it will remain merely an internal experimental tool unavailable to users. An engineer reported that early versions of the Argon model suffered from slowness and a regression in certain programming tasks compared to Anthropic’s older Claude Opus 5 model.
Details of the Argon model and the Fairwind cybersecurity program
Google had officially announced the launch of the “Gemini 4 Argon” model on September 30, limiting its initial distribution to an elite group of trusted experts and cyber defenders engaged in the “Fairwind program.” The company stated that it would expand the availability of the model as soon as field feedback is gathered and digital safety guardrails are strengthened.
The company confirmed that Argon achieved a new record of 77.9 percent on the “DeepSWE 1.1” test, a rigorous and prolonged evaluation to measure software engineering proficiency. The company also raised the model’s output limit to one million tokens per response compared to 64,000 tokens previously. Introductory pricing was set at $2 per million input tokens and $10 per million output tokens. The model is expected to be made available later to paid API subscribers and the AI Ultra plan, after technical platforms detected signs of its use in the “Antigravity” programming tool.
Internal debate at Google regarding model efficiency and technical leadership
The day of the Argon announcement was accompanied by a report published by Bloomberg indicating internal skepticism among some Alphabet personnel regarding the model’s performance efficiency in real-world work tasks; some specialists who tested the system directly felt that despite scoring high marks on synthetic benchmarks, its performance in actual field use, particularly in user interface design, was less convincing, noting that competing companies are developing their models at a faster pace.
In response to these concerns, Google asserted that “it is inaccurate to say that Gemini 4 is subpar in areas such as coding.” An employee close to the development team stated that there is a broad consensus within the company that “Gemini 4 stands at the forefront of advanced technology.” Koray Kavukcuoglu, head of research at Google DeepMind, also emphasized this direction, stating: “In my view, it is absolute certainty that we will always remain at the forefront and leadership of this fast-paced field.”
