Google launches Gemini 4 Carbon to the public with functionality for AI agent teams

In June, Google extended a voluntary redundancy program to employees located in Canada, France, Germany, Japan, and Switzerland. This action is part of Alphabet's efforts to control costs amid intensifying investments in artificial intelligence (AI).

The company confirmed this expansion to Reuters, although it did not disclose the exact number of employees who accepted the separation offers. This program follows similar initiatives offered since 2025 to employees in certain units in the United States and the United Kingdom. A spokesperson mentioned that in that year, some teams accepted voluntary exit packages with compensation before this international expansion in June 2026.

Alphabet seeks to concentrate its resources on sectors considered strategic, even while recording an increase in staff in recent quarters. Chief Financial Officer Anat Ashkenazi emphasized artificial intelligence and cloud computing as priority areas for new hires during quarterly earnings presentations.

In previous offers, employees of Google Cloud and DeepMind, the company's AI division, were excluded from the programs. The most recent expansion appears to maintain this same pattern, according to Reuters. Pressure to manage expenses is also observed in other industry corporations; for example, Microsoft offered voluntary redundancy programs to 60,000 employees the previous week.

Investments in AI have caused apprehension among Wall Street investors and exerted pressure on Alphabet's shares. In Switzerland, where Google has about 5,000 employees and operates a significant engineering center in Zurich, the program reached several divisions but maintained the exclusion of DeepMind, according to sources familiar with the matter spoken to by Reuters.

The company released the proposals in mid-June and set the deadline for interested employee responses until mid-July, with scheduled departures beginning in August. The specific conditions offered in Switzerland include the fact that this episode occurs three years after a previous staff cut in Switzerland, when Google dismissed more than 200 employees, which resulted in protests and worker strikes.

Two weeks after the launch of Gemini 4 Argon, Google introduced Gemini 4 Carbon, which the company considers its new flagship model. This model was specifically developed to handle advanced programming and cybersecurity activities, including identifying flaws in controlled environments and correcting code.

The model was presented this Wednesday, October 29th, but it is not yet available to the general public. Following the precedent set with Claude Mythos 5 for large-scale models, Carbon will initially be available only to digital defense specialists while Google conducts security assessments with United States authorities.

Google plans to release initial API access within one to two weeks. However, there is no defined timeline for its integration into the Gemini application or the browser interface. API costs are expected to follow the same structure as the Argon model: US$ 2 per million input tokens and US$ 10 per million output tokens.

Regarding performance results, Google highlights the progress of artificial intelligence in benchmarks. In the independent consultancy Artificial Analysis index, which evaluates AI performance in near real-world scenarios, Gemini 4 Carbon achieved a score of 59 points. This score places it ahead of competitors such as Claude Opus 5.5 (Anthropic), which scored 58, GPT-6 Astra (OpenAI) with 57, and Gemini 4 Argon itself, with 50.

According to Google DeepMind, one of the most significant advances in Carbon lies in its ability to manage extensive tasks. In internal tests, the model demonstrated the ability to work on complex software problems for several days, successfully resuming processes and seeking alternative solutions when encountering dead ends.

Despite the advancements, areas were identified where Carbon performed below its rivals. Specifically in the FrontierSWE 2 benchmark, focused on programming agents, it lagged behind GPT-6 Astra and Opus 5.5. Furthermore, Anthropic's AI surpassed Carbon in the OSWorld-2.1 test, one of the main indicators of autonomous operational capability of computer systems.

The primary focus of Gemini 4 Carbon is the cybersecurity sector. In tests conducted by the company, the model demonstrated the capacity to act from vulnerability discovery to the implementation of possible fixes. It is relevant to note that this focus comes after data leakage incidents in test environments involving models from OpenAI, Anthropic, and Meta.

Additionally, Carbon exhibited high resistance against prompt injection attacks, an issue that has been particularly concerning for companies and government agencies that delegate tasks to AI. In all tests conducted, the Google model resisted all attempts at manipulation through prompt injection.

General Motors gradually began making Gemini, the artificial intelligence assistant developed by Google, available for Chevrolet vehicles in Brazil that have the integrated Google system, this Wednesday (28th).

Initially, the technology was launched in English in the United States in April, subsequently expanding to Spanish-speaking Latin American countries, and is now arriving to Brazilian drivers in Portuguese.

This system differs from conventional voice commands by understanding natural language and maintaining the conversation context across multiple interactions, eliminating the need for the driver to repeat the initial command.

In practice, Gemini allows actions such as adjusting the temperature by saying 'I'm cold' or activating the defroster by stating that the windshield is foggy. Additionally, it is possible to request suggestions for stops during the trip or search for music without knowing the exact title.

The assistant also has the ability to consult the owner's manual to clarify doubts about dashboard lights. In electric models, it provides information about battery range to the destination and plans routes with access to charging points.

Through Gemini Live, activated by the phrase 'Hey Google, let's chat,' drivers and passengers can maintain a continuous dialogue to organize itineraries. Users have the option to return to the traditional Google Assistant through the settings menu.

This update is available for Chevrolet Blazer EV, Equinox EV, Equinox Turbo, and Silverado models manufactured from 2024 onwards, distributed via the Google Play store.

For the functionality to be activated, the vehicle must run Android 11 or a higher version, possess native internet connectivity, have a Google account linked to the multimedia center, and an active OnStar plan, which is Chevrolet's safety and connectivity service.

In models planned for 2027, Gemini will be offered free of charge with the OnStar Basics package. However, for 2026 models and earlier years, the service subscription remains a paid requirement.

Access to the assistant can be made via voice command, using the button located on the steering wheel or by tapping the icon displayed on the central screen, which reduces the need for the driver to navigate through various menus while driving.