09:34Machine-translated from Azerbaijani by AIRead in Azerbaijani
Google Unveils New Gemini 4 Carbon AI Model, Leads in 15 of 17 Tests
Google announced the Gemini 4 Carbon artificial intelligence model on October 28, 2026. The company calls it the most powerful model it has developed to date. Carbon was unveiled just 2 weeks after Gemini 4 Argon became available. According to Google, the model excels at cybersecurity and long, multistage tasks. It can work on complex programming tasks for days.
According to Google’s comparison, Gemini 4 Carbon achieved the highest result in 15 of 17 tests. The model took first place in the Artificial Analysis Intelligence Index with 59 points. It surpassed Gemini 4 Argon, which scored 50 points. However, GPT-6 Astra remained the leader in FrontierSWE 2, a test that measures software engineering. Claude Opus 5.5 achieved the best result in OSWorld-2.1, which tests the ability to operate a computer. In the independent Terminal-Bench 4.0 test, Carbon scored about 8 points more than the previous model. This test evaluates the use of terminal commands to perform real-world tasks such as installing software and fixing faulty code. The model rose 89 points to 4th place on the WebDev Arena leaderboard. It achieved a result of 88.2 percent in the SEC-bench Pro cybersecurity test. Its result was 54.8 percent in ExploitGym and 71 percent in CWE-bench. In Google’s internal tests, no successful attempt to achieve a good result without completing the task was recorded. The overall rate of attempts to seek such shortcuts was just 0.4 percent. Alerts related to serious misaligned behavior fell by about 80 percent. The number of unfounded claims was also cut in half. According to the company, all attacks that attempted to steer the model with hidden instructions were blocked. With Argon, 0.7 percent of such attacks had been successful.
Gemini 4 Carbon is currently available only to selected cybersecurity specialists through the Google Fairwind Program. Starting this week, Google will gradually increase the number of trusted partners. Developers, companies and regular users will be able to use the model later, but no exact date has been announced. The company is also working with the US government to test the model.
When the API—the system for connecting the model to other applications—is released, the initial price will be 2 dollars (about 3 manats) per 1 million input tokens. A token is a small piece of text processed by artificial intelligence. Every 1 million output tokens will cost 10 dollars (17 manats). Previously cached and reused input tokens will receive a 95 percent discount. The input price is expected to rise later to 4 dollars (about 7 manats). The price of output tokens may also increase to 20 dollars (34 manats). For comparison, GPT-6 Astra costs 10 dollars (17 manats) for input tokens and 50 dollars (85 manats) for output tokens. Google also unveiled the Genie 4 model this week. This world model was developed to simulate real and imagined digital environments in accordance with the laws of physics. The model, which can be used in robotics and video game development, also generates background sounds and dialogue. Genie 4 is not yet widely available, but users can join the waitlist.
According to Google’s comparison, Gemini 4 Carbon achieved the highest result in 15 of 17 tests. The model took first place in the Artificial Analysis Intelligence Index with 59 points. It surpassed Gemini 4 Argon, which scored 50 points. However, GPT-6 Astra remained the leader in FrontierSWE 2, a test that measures software engineering. Claude Opus 5.5 achieved the best result in OSWorld-2.1, which tests the ability to operate a computer. In the independent Terminal-Bench 4.0 test, Carbon scored about 8 points more than the previous model. This test evaluates the use of terminal commands to perform real-world tasks such as installing software and fixing faulty code. The model rose 89 points to 4th place on the WebDev Arena leaderboard. It achieved a result of 88.2 percent in the SEC-bench Pro cybersecurity test. Its result was 54.8 percent in ExploitGym and 71 percent in CWE-bench. In Google’s internal tests, no successful attempt to achieve a good result without completing the task was recorded. The overall rate of attempts to seek such shortcuts was just 0.4 percent. Alerts related to serious misaligned behavior fell by about 80 percent. The number of unfounded claims was also cut in half. According to the company, all attacks that attempted to steer the model with hidden instructions were blocked. With Argon, 0.7 percent of such attacks had been successful. Gemini 4 Carbon is currently available only to selected cybersecurity specialists through the Google Fairwind Program. Starting this week, Google will gradually increase the number of trusted partners. Developers, companies and regular users will be able to use the model later, but no exact date has been announced. The company is also working with the US government to test the model. When the API—the system for connecting the model to other applications—is released, the initial price will be 2 dollars (about 3 manats) per 1 million input tokens. A token is a small piece of text processed by artificial intelligence. Every 1 million output tokens will cost 10 dollars (17 manats). Previously cached and reused input tokens will receive a 95 percent discount. The input price is expected to rise later to 4 dollars (about 7 manats). The price of output tokens may also increase to 20 dollars (34 manats). For comparison, GPT-6 Astra costs 10 dollars (17 manats) for input tokens and 50 dollars (85 manats) for output tokens. Google also unveiled the Genie 4 model this week. This world model was developed to simulate real and imagined digital environments in accordance with the laws of physics. The model, which can be used in robotics and video game development, also generates background sounds and dialogue. Genie 4 is not yet widely available, but users can join the waitlist.