Agnes 2.5 Pro Beta

Agnes 2.5 Pro Beta is a large language model developed by Sapiens AI, a Singapore-based research laboratory. Released in August 2026, it serves as a high-performance reasoning model optimized for professional-tier workloads, specifically targeting agentic tasks, complex software engineering, and large-scale scientific data analysis. Compared to its Alpha predecessor, the Beta release demonstrates a significant jump in multi-step workflow efficiency and tool-calling accuracy.

The model is built on a sparse Mixture-of-Experts (MoE) architecture, which allows it to maintain high inference throughput while possessing a high total parameter count, reportedly reaching 1 trillion parameters. It features a 1 million token context window, enabling the analysis of extensive codebases and long-form technical documentation. On independent benchmarks like the Artificial Analysis Intelligence Index, Agnes 2.5 Pro Beta has been recognized for its cost-performance, ranking near the Pareto frontier for reasoning capabilities relative to its API pricing.

Capabilities and Performance

Agnes 2.5 Pro Beta is designed for long-horizon reasoning, where the system must autonomously plan and execute a series of connected steps. Key performance metrics show a substantial reduction in hallucination rates—dropping significantly from earlier iterations—and strong proficiency in specialized benchmarks such as Terminal-Bench and scientific reasoning tests. The model supports multimodal inputs, allowing users to process combined text and image data for architecture-level debugging and visual data interpretation.

For best results, users are encouraged to utilize structured prompts when handling long-context workloads, such as including explicit file markers or section headers to help the model maintain traceability. When performing complex engineering tasks, providing a high-level summary of the target architecture alongside specific error logs typically maximizes the model's patch-generation accuracy. The model supports an extended output length of up to 65,536 tokens, making it suitable for generating full application modules or exhaustive technical reports.