Gemini 4 Carbon Leads, but the Race Remains Open

Google unveiled the Gemini 4 Carbon artificial intelligence model on 28 October 2026. According to the company’s official announcement, the new system primarily promises advances in software development and cybersecurity tasks.

Carbon has taken the lead in one composite model ranking, but this alone does not mean it outperforms its competitors across every type of task. For example, Claude Opus 5.5 remained ahead in a comparison focused specifically on coding agents. For developers and businesses, the more important question is therefore not the ranking, but what results, costs and risks the model brings to their own tasks.

Rapid turnover among Google’s flagship models

According to Engadget’s report, only 15 days passed between the announcements of Gemini 4 Argon and Carbon. This short interval illustrates how quickly new flagship models can follow one another, although it does not reveal how much the development cycles of the two systems overlapped.

Google has not disclosed Carbon’s parameter count or detailed internal architecture. For now, the new model can therefore be assessed primarily through its published benchmark results, access conditions and experience gained from practical use.

What does first place actually mean?

According to Scaler’s comparison, Gemini 4 Carbon scored 59 points in the cited composite ranking. In the same comparison, Claude Opus 5.5 received 58 points and GPT-6 Astra 57. The margin is therefore narrow, and the result reflects a specific measurement method.

Benchmarks can combine several capabilities into a single score, so first place overall should not be interpreted as universal superiority. One model may excel at debugging or security analysis, while another performs better on longer programming tasks requiring tool use.

This is also supported by a technical analysis published on dev.to: according to the article, Claude Opus 5.5 scored 62 points in an index focused on coding agents, while Gemini 4 Carbon scored 59. The measurements therefore indicate that Carbon’s overall performance is outstanding, but it does not win the coding category in every environment.

Cybersecurity is Carbon’s most important field

According to Google, Gemini 4 Carbon scored 88.2 per cent on SEC-bench Pro, which measures vulnerability detection. The company reported a score of 54.8 per cent on the ExploitGym exploitability assessment and 71 per cent on CWE-bench, which measures vulnerability repair.

These figures suggest that the model could be useful at several stages of identifying security flaws: it may detect a problem, analyse its exploitability and then suggest a fix. However, this does not remove the need for human review, as an automatically generated fix may introduce a new flaw or alter how the software works.

According to The Hacker News, Carbon was twice as likely as Argon to produce secure code in Google’s internal tests. As this is a manufacturer’s measurement, independent assessments and experience in real-world environments may provide a more accurate picture of its practical significance.

Access is currently limited to a small group

According to Google, cybersecurity experts in the Fairwind Program will have access to the model during the first phase of its rollout. Based on SecurityWeek’s report, more than 650 organisations participate in the programme, including government institutions, critical infrastructure operators and technology companies.

The limited launch reflects the model’s dual-use capabilities. A system that can identify and repair vulnerabilities may support defensive work, but the same capabilities can also be used for offensive purposes. Gradual access is therefore not merely a product launch decision, but also a security precaution.

A public API is not yet available to the general public or the wider developer community, and Google has not provided a precise date for broader access. This means that most businesses cannot yet build a firm deployment timetable around Carbon.

Pricing may look favourable, but it is not the full cost

According to Google’s announcement, the introductory Fairwind Program price is $2 per million input tokens and $10 per million output tokens. The standard pricing published by the company may later be $4 and $20 respectively per million tokens.

Token pricing, however, is only one cost within an enterprise system. Integration, access management, logging, data protection, result verification and the consequences of incorrect responses must also be considered. A seemingly inexpensive model can become costly if it produces overly long answers, requires many repeated runs or frequently needs human correction.

It is therefore worth testing the system on a narrow, clearly measurable task. This could include supporting code review or searching for a specific type of vulnerability. The comparison should consider not only accuracy, but also total running costs, response times and the work required for human review.

What should developers and businesses consider?

Carbon’s announcement has three important implications. First, it further strengthens the trend of flagship models competing increasingly in specific workflows, such as coding and security review, alongside general text generation. Second, rapid model turnover may make it risky to build an entire system around the characteristics of a single provider.

Third, benchmark results must be supplemented with local testing. An organisation’s codebase, programming languages, documentation quality and security policies may differ significantly from the environments used in public tests.

• Task-specific measurement: the same real-world workflow should be tested with several models.
• Assessing total cost: integration and review expenses should be considered alongside token costs.
• Secure isolation: access permissions, logging and data-processing terms must be clarified before sensitive code is processed.
• Human approval: automatically suggested fixes need to be tested and professionally reviewed.
• Ability to switch providers: it may be beneficial to design an architecture that is not unnecessarily tied to a single model.

A strong start, but not a definitive change at the top

Gemini 4 Carbon’s results show that Google has returned to the leading group and taken first place in the cited composite ranking. Its cybersecurity results are particularly promising, while the rapid release schedule indicates that the order among flagship models can change within a short period.

However, the winner of the “model war” cannot be determined from a single score. Claude Opus 5.5 continued to appear stronger in certain coding-agent tasks, while Carbon still lacks broad API access. For developers and businesses, controlled testing on their own tasks is therefore a better guide than brand loyalty.