MiniMax M2.5 Open-Source Model Challenges Claude Opus 4.6 in Performance and Cost

MiniMax has launched its M2.5 open-source model, which the company claims rivals the coding performance of Claude Opus 4.6 while offering a significantly lower price point. The model is available at $1 per hour, with MiniMax stating that this cost-effectiveness allows for sustained high performance even with limited computing resources. Following the announcement, MiniMax's market value reportedly surged by 16.2%, reaching a historical high of HKD 210.8 billion.

Developer's hands typing code on a laptop, with performance metrics on screen.

The M2.5 model has reportedly closed the performance gap with leading large language models in critical coding benchmarks. MiniMax asserts that M2.5 is the first open-source model to surpass Claude Sonnet in performance.

Technical Capabilities and Cost Efficiency

MiniMax attributes M2.5's capabilities to an agent-native architecture optimized for the agent ecosystem and extreme inference efficiency. The model supports a throughput of 100 tokens per second (TPS) and an inference speed three times faster than Claude Opus 4.6, according to MiniMax. The company states that M2.5 can run for one hour at 100 TPS for $1.

Glowing, compact microchip at the center of a digital network, symbolizing efficiency.

The model also reportedly excels in core productivity scenarios such as Excel, deep search, and long document summarization. With an activation parameter count of 10 billion, M2.5 is positioned as the smallest flagship model in its tier, offering advantages in private deployment, video memory footprint, and inference energy efficiency. MiniMax has expanded its capacity to accommodate demand for M2.5.

Performance Across Development and Office Tasks

Tests conducted by toolmesh.ai indicate that MiniMax M2.5 demonstrates high adaptability across various developer tools, including IDEs like Claude Code, Trae, and VS Code. It integrates into existing development workflows for code generation, debugging assistance, and multi-turn context understanding.

In a "car wash" logic problem, M2.5 provided a correct answer in three seconds, while Claude Opus 4.6 reportedly failed a variant of the same problem. For a LeetCode Hard algorithm problem, M2.5 delivered a solution with optimal time complexity in 20 seconds, outperforming Claude Opus 4.6, which provided a solution with higher time and space complexity.

The model also showed proficiency in complex coding tasks, including code review, refactoring, and architectural design on existing codebases. It identified non-atomic operations in a Java Counter class and analyzed missing error handling in a Python function.

Person analyzing complex Excel data on a monitor, showing data cleaning and analysis.

Beyond coding, MiniMax M2.5 demonstrated strong capabilities in digital office scenarios. When presented with a complex Excel spreadsheet containing dirty sales data, the model successfully cleaned the data, performed cross-table associations, and conducted analyses, including identifying the product category with the lowest gross profit margin. It also accurately generated a list of Forbes Global Billionaires from 2019 to 2024 in an Excel table format.

Advanced Research and Long Text Processing

MiniMax M2.5's research capabilities were tested by asking it to conduct an in-depth study of OpenClaw. The model generated a detailed and logically structured research report, demonstrating an understanding of core concepts, technical details, and an ability to evaluate the subject from multiple perspectives.

For long text processing, M2.5 was tasked with interpreting "Principles of Mathematical Analysis" by Rudin. The model extracted core concepts, provided intuitive analogies, and identified common misconceptions for each chapter. While minor errors were noted, the model effectively summarized complex mathematical content and conveyed core intuitions, according to toolmesh.ai's assessment.

Holographic projection of research data and equations above an open book, symbolizing advanced text processing.

Underlying Technical Breakthroughs

MiniMax attributes M2.5's performance to its Agent Reinforcement Learning (Agent RL) technical system. The team developed Forge, a native Agent RL framework designed to decouple the underlying training and inference engine from the agent, supporting the integration of various agents and optimizing generalization. This framework reportedly achieved a 40-fold acceleration in training.

The team utilized the CISPO algorithm, previously introduced with MiniMax M1, to ensure stability in large-scale reinforcement learning training for Mixture of Experts (MoE) models. To address credit assignment in long-context agent scenarios, MiniMax introduced a process reward mechanism to monitor completion quality. The system also estimates the actual task execution time in real environments as a reward function, balancing model effectiveness and response speed.

Accelerating, interconnected glowing nodes in a neural network, symbolizing rapid AI training.

MiniMax views M2.5's performance as an indicator that top-tier AI capabilities are becoming more accessible. The company suggests that future differentiation in the industry will depend on system integration, productization, and understanding of real-world scenarios.