Google's Gemini 4 Carbon takes top spot in AI benchmarks

A return to the top

Google recently released its new flagship model, Gemini 4 Carbon, just before Alphabet's Q3 earnings call. The model quickly secured the top spot on the Artificial Analysis Intelligence Index, earning a score of 59. This marks a return to the number one position for Google after about half a year of being surpassed by rivals.

The new score puts Carbon ahead of major competitors like Claude Opus 5.5 and GPT-6 Astra. Specifically, Carbon outperformed Argon, the model Google opened to paying users just two weeks ago, by 9 points. It showed particular strength in complex reasoning, scientific research, and automated tasks.

Despite this technical milestone, the company's financial report was less positive. Alphabet reported a Q3 free cash flow of negative $9.065 billion, leading to a stock price drop of more than 10% in after-hours trading. This creates a stark contrast between the model's high benchmark performance and the company's current financial health.

Coding and security gains

One of the most notable improvements for Gemini 4 Carbon is in its programming capabilities. In the Coding Agent index, it scored 59 points, trailing only Claude Opus 5.5. This is a significant jump for Google, which has historically struggled to lead in these specific types of programming tests.

The model also showed marked progress in terminal programming and web development. In the Terminal-Bench 4.0 test, Carbon's score rose from 57.1% for Argon to 65.2%. On the WebDev Arena, it saw an increase of 89 Elo points, placing it fourth overall with 1768 points as recorded on the original leaderboard.

Security is another area where Carbon excels. It successfully resisted every attempt in the Gray Swan indirect prompt injection test, which measures an AI's ability to withstand malicious instructions hidden in external materials. In comparison, other models like GPT-6 Astra had an 8.5% attack success rate in the same test.

Long-running agent workflows

Google designed Carbon to handle long-running agent workflows, which are tasks that require the AI to work continuously over time. Internal tests show the model can run for several days to locate and fix complex bugs in large codebases. This suggests it can take over tasks that previously required many hours of human engineering involvement.

To ensure reliability, Google claims Carbon is an "honest model." The company reports that the model's cheating success rate in the Terminal-Bench 4.0 test dropped to 0%. Furthermore, the count of statements made by Carbon that lacked supporting evidence was reduced by about half when compared to the previous Argon model.

While the model is more proactive and less likely to "go through the motions," it faces some hurdles in accuracy. Some tests show that while knowledge Q&A accuracy increased from 48% to 58%, the hallucination rate—where a model gives a wrong answer instead of admitting it doesn't know—rose from 4% to 11%.

Tradeoffs and availability

Early access testing reveals that Carbon's increased power comes with a significant sacrifice in speed. Some developers reported that Carbon is 2 to 4 times slower than Argon depending on the task. For example, in legal research tests, Carbon took nearly 52 minutes on average per task, which was about 20 minutes longer than the previous model.

Pricing for Carbon currently mirrors Argon's preferential rates of $2 per million input tokens and $10 per million output tokens. Once this period ends, prices are scheduled to rise to $4 and $20 respectively. Even at higher prices, Carbon's cost to complete a comprehensive evaluation task remains relatively competitive compared to Claude Opus 5.5.

However, the general public cannot use Carbon just yet. Google is currently rolling out the model to a small number of cybersecurity defense institutions and trusted partners through the Fairwind program. There is currently no official timeline for when ordinary consumers or enterprise customers will gain full access to the new flagship.