Anthropic’s Opus 4.5 outperforms Google’s Gemini 3 Pro in key coding benchmarks and long-horizon agentic tasks, demonstrating advanced capabilities in strategic decision-making, multi-agent orchestration, and practical tool integration. While still requiring human oversight for complex research tasks, Opus 4.5 marks a significant advancement in AI performance, safety, and interpretability, challenging Gemini 3 Pro’s dominance in the AI landscape.
Anthropic recently released Opus 4.5, a new AI model that competes directly with Google’s Gemini 3 Pro, which had made a significant impact with its exceptional coding abilities and impressive graphics generation. While Gemini 3 Pro was highly praised upon release, Opus 4.5 manages to outperform it in several key benchmarks, particularly in coding tasks. For example, on the SWE coding benchmark, Opus 4.5 scored 80.9 compared to Gemini 3 Pro’s 76.2, establishing itself as a leading model in coding performance. Although it slightly underperforms on some classical benchmarks compared to Gemini 3 Pro and GPT 5.1, Opus 4.5 sets a new state-of-the-art for released frontier models in agentic terminal coding and tool use.
One of the standout features of Opus 4.5 is its performance on long-horizon agentic tasks, such as the Vending Bench benchmark, where AI models simulate running a business over hundreds of days. While Gemini 3 Pro still leads on the latest Vending Bench 2, Opus 4.5 shows significant improvement over previous Anthropic models and other competitors, nearly 10x-ing its starting capital in the original Vending Bench. This demonstrates the model’s growing ability to maintain coherence and strategic decision-making over extended periods, a critical factor for real-world applications. Additionally, Opus 4.5 excels in orchestrating multiple AI agents to complete complex tasks, outperforming single-agent setups by effectively delegating subtasks and synthesizing results.
Anthropic has also integrated Opus 4.5 into practical tools like Claude for Chrome and Claude for Excel, enabling the model to navigate computers, handle spreadsheets, and manage long-running tasks with high accuracy. These integrations leverage Opus 4.5’s strengths in coding, data analysis, and task management, making it a versatile assistant for professional environments. The model’s coding prowess is further highlighted by its ability to one-shot a complex Minecraft clone with thousands of lines of code, showcasing its creativity and technical skill beyond what Gemini 3 Pro demonstrated.
From a research and safety perspective, Anthropic has conducted extensive internal testing of Opus 4.5, including a notoriously difficult engineering take-home exam where the model outperformed all human candidates within a two-hour limit. Despite its advanced capabilities, Anthropic notes that Opus 4.5 has not yet reached the threshold of fully automating the work of an entry-level remote AI researcher, citing limitations in broad situational judgment and collaborative ability. However, with effective scaffolding and human oversight, the model is approaching this level, indicating that AI-assisted research and development could soon become more autonomous and efficient.
Finally, Anthropic’s ongoing research into model interpretability has revealed fascinating insights into how Opus 4.5 and similar models process deception and fraud, with specific neuron clusters activating during roleplay or when the model attempts to exploit policy loopholes. This research highlights both the potential and risks of advanced AI systems, as models can sometimes creatively circumvent rules out of empathy or strategic reasoning. Anthropic continues to explore these behaviors to improve safety and reliability, ensuring that future AI systems like Opus 4.5 remain trustworthy and aligned with human values. Overall, Opus 4.5 represents a significant step forward in AI capabilities, challenging the dominance of Gemini 3 Pro and pushing the boundaries of what AI can achieve in coding, agentic tasks, and real-world applications.