# Minimax M2.5 - What Makes This Different!

Source: https://www.youtube.com/watch?v=f1DzkFc9vxo
Recap page: https://rapidrecap.app/video/f1DzkFc9vxo
Generated: 2026-02-13T12:04:01.633+00:00

---
## Quick Overview

MiniMax M2.5 achieves state-of-the-art (SOTA) performance in coding and agentic tasks, matching or exceeding proprietary models like Claude Opus 4.6 and GPT-5.2 Codex on several benchmarks, while offering significantly lower operational costs, exemplified by a $1/hour continuous runtime at 100 tokens per second output.

**Key Points:**
- MiniMax M2.5 achieves SOTA performance in coding benchmarks, boasting 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp (with context management).
- The M2.5 model demonstrates tremendous speed, completing the SWE-Bench Verified evaluation 37% faster than MiniMax M2.1, matching the speed of Claude Opus 4.6.
- The new M2.5 model is significantly cost-effective, costing $1 per hour to run continuously at 100 output tokens per second, or $0.30 per million input tokens and $2.4 per million output tokens for the M2.5-Lightning variant.
- M2.5's output cost is one-tenth to one-twentieth that of top proprietary models like Opus, Gemini 3 Pro, and GPT-5.
- MiniMax has developed an agent-native RL framework called Forge, which decouples training/inference, supports arbitrary agents, and achieves approximately 40x training speedup using tree-structured merging strategy.
- The company has built over 10,000 Experts on MiniMax Agent, with 30% of daily tasks autonomously completed by M2.5, showing strong performance in coding scenarios (80% of newly committed code).
- The OpenRouter pricing shows MiniMax M2.5 (fp16) costing $0.30/$1.20 input/output per million tokens, significantly cheaper than competitors like Claude Opus 4-6 ($0.30/$10.00) and GPT-5.2 Codex ($0.30/$75.00) on input price, and far cheaper on output price.

![Screenshot at 00:08: The pricing table comparison for Claude models highlights the significant cost differences between various tiers, setting the stage for MiniMax M2.5's cost advantage.](https://ss.rapidrecap.app/screens/f1DzkFc9vxo/00-00-08.jpg)

**Context:** The video discusses the release and capabilities of MiniMax M2.5, a new frontier large language model from the Chinese AI company MiniMax, focusing on its performance in real-world productivity and agentic tasks, particularly coding. The presentation contrasts M2.5's performance and extreme cost-effectiveness against established proprietary models like Anthropic's Claude family and OpenAI's GPT series, highlighting the company's underlying reinforcement learning (RL) infrastructure called Forge.

## Detailed Analysis

MiniMax officially announces its latest frontier model, MiniMax M2.5, emphasizing its capabilities in real-world productivity and agentic tasks, particularly coding, where it achieves state-of-the-art (SOTA) results. The model scored 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp. Furthermore, M2.5 exhibits significant speed improvements, being 37% faster than its predecessor M2.1 and matching the speed of Claude Opus 4.6. A major focus is cost-effectiveness: running M2.5 continuously for an hour at 100 tokens per second costs only $1. In terms of token pricing, the M2.5-Lightning version costs $0.3 per million input tokens and $2.4 per million output tokens, while the standard M2.5 costs one-tenth to one-twentieth of proprietary models like Opus, Gemini 3 Pro, and GPT-5 on output price. The video also details the underlying technology, Forge, an agent-native RL framework designed to decouple training-inference engines and optimize model generalization across scaffolds, achieving a 40x training speedup through asynchronous scheduling and tree-structured merging. This infrastructure is crucial for training the model iteratively across thousands of synthetic RL environments. The OpenRouter pricing screenshot confirms the cost advantage, showing MiniMax M2.5 input price at $0.30/M tokens compared to $0.30/M for others but significantly cheaper output prices ($1.20 vs. $10.00 for Claude Opus 4-6 and $75.00 for GPT-5.2 Codex).

### Model Performance Benchmarks

- M2.5 achieves SOTA scores: 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp
- M2.5 is 37% faster than M2.1 on SWE-Bench Verified, matching Claude Opus 4.6 speed

### Cost Efficiency

- Continuous runtime at 100 tokens/sec costs $1/hour; Output price is 1/10th to 1/20th of proprietary models like Opus and GPT-5
- M2.5-Lightning costs $0.3/M input and $2.4/M output tokens

### Agentic RL Framework (Forge)

- Designed as an agent-native RL framework using an intermediary layer to decouple inference engine from agent
- Optimizes model generalization across scaffolds and tools using asynchronous scheduling and tree-structured merging strategy

### Training Scale

- MiniMax has made hundreds of thousands of environments for RL training
- Agents are trained using on-policy and off-policy data collected from these environments

### MiniMax Agent Deployment

- M2.5 is deployed in MiniMax Agent, handling standardized Office Skills (Word, PowerPoint, Excel)
- Over 10,000 Experts built, with 30% of daily tasks autonomously completed by M2.5

### OpenRouter Pricing Comparison

- MiniMax M2.5 (fp16) costs $0.30/$1.20 input/output per million tokens, significantly undercutting competitors like Claude Opus 4-6 ($0.30/$10.00) and GPT-5.2 Codex ($0.30/$75.00) on output price.

![Screenshot at 00:08: The pricing table comparison for Claude models highlights the significant cost differences between various tiers, setting the stage for MiniMax M2.5's cost advantage.](https://ss.rapidrecap.app/screens/f1DzkFc9vxo/00-00-08.jpg)
![Screenshot at 00:48: Bar charts comparing MiniMax M2.5 performance against competitors across various coding benchmarks like SWE-Bench Verified and BFCL multi-turn.](https://ss.rapidrecap.app/screens/f1DzkFc9vxo/00-00-48.jpg)
![Screenshot at 01:01: Appendix table showing detailed benchmark scores for M2.5 across several technical evaluations \(AIME25, GPQA-D, etc.\) compared to other models.](https://ss.rapidrecap.app/screens/f1DzkFc9vxo/00-01-01.jpg)
![Screenshot at 02:20: The MiniMax M2.5 landing page banner emphasizing speed, strength, and smartness for real-world productivity.](https://ss.rapidrecap.app/screens/f1DzkFc9vxo/00-02-20.jpg)
![Screenshot at 08:23: Diagram illustrating the Forge framework, separating the Agent \(Black Box/White Box\) from the Middleware and Engines, showing data flow for RL training.](https://ss.rapidrecap.app/screens/f1DzkFc9vxo/00-08-23.jpg)