Gemini 4 Carbon Tops Vending Bench, First Google Model To Top Benchmark In Over A Year

Google's Gemini 4 Carbon finished Vending-Bench 2 with a leading bank balance of $16,484, topping the Andon Labs benchmark ahead of GPT-6 Astra and Claude Opus 5.5 and marking the first time in over a year a Google model has taken the #1 spot on it. The simulation gives each model a starting balance such as $500 and tasks it with running a vending machine business across 365 simulated days, paying daily location fees, sourcing inventory from competing suppliers, handling delivery delays and adjusting pricing across 3,000 to 6,000 messages and tens of millions of tokens. The result follows Gemini 4 Carbon's earlier top placement on the Artificial Analysis Intelligence Index and signals Google's progress on long-horizon agentic consistency, an area where its models had lagged Anthropic's Claude series for over a year.

Google appears to have made it out of the AI wilderness with Gemini 4 Carbon. After topping https://officechai.com/ai/googles-gemini-4-carbon-tops-artificial-analysis-intelligence-index-becomes-top-ai-model-in-the-world/ the Artificial Analysis Intelligence Index after nearly a year, Google has claimed the top spot on Vending-Bench 2 , marking the first time in over a year that a Google model has clinched the 1 position on this grueling, real-world operational benchmark. Surpassing rival models like GPT-6 Astra and Claude Opus 5.5, Gemini 4 Carbon finished the simulation with a leading bank balance of $16,484 , showcasing a massive leap forward in long-horizon reasoning and financial execution. What Is Vending-Bench and How Does It Work? Developed by Andon Labs, Vending-Bench 2 is widely considered one of the toughest stress tests for autonomous AI agents. Unlike traditional benchmarks that test short-form, single-turn knowledge such as answering a multiple-choice question or writing a single function , Vending-Bench evaluates how well an AI model can manage an actual business over a long time horizon . Here is how the simulation operates: - The Core Task: Each model is given a starting balance such as $500 and is tasked with operating a vending machine business across 365 simulated days . - Operational Constraints: The AI agent must pay a daily location fee, source and purchase inventory from multiple competing suppliers, handle delivery delays, navigate supplier negotiations, and adjust product pricing. - State Management & Consistency: Customer sales fluctuate based on real-world variables like days of the week, seasonality, pricing, and weather patterns. Because the simulation spans thousands of interactions generating between 3000 to 6000 messages and consuming tens of millions of tokens , the model cannot rely on isolated logic. It must track its inventory state, remember past decisions, avoid strategic drift, and recover from mistakes without losing its operational rhythm. Why Gemini 4 Carbon’s Win Matters For Google, topping Vending-Bench 2 represents a critical turning point. For over a year, Google models lagged behind competitors like Anthropic’s Claude series in long-horizon consistency and multi-step execution. Gemini 4 Carbon’s record-breaking $16,484 final bank balance demonstrates that Google has successfully solved key agentic hurdles—specifically, maintaining long-term context without degrading performance, optimizing supplier negotiation strategies, and executing compounding financial decisions over months of simulated time. As businesses increasingly look to deploy autonomous AI agents capable of managing complex, real-world workflows rather than simple chat prompts, Gemini 4 Carbon’s dominant performance on Vending-Bench 2 signals that Google is ready to compete at the absolute frontier of enterprise automation.