Chinese AI startup MiniMax has launched its M2.5 model, marking a significant leap in cost efficiency for autonomous AI agents while achieving state-of-the-art performance across programming, tool integration, and productivity applications. The Shanghai-based company claims the model can run complex agentic tasks continuously for one hour at just $1, potentially removing economic barriers to widespread AI agent deployment.

The release comes 108 days after the company's initial M2 launch, representing what MiniMax describes as the industry's fastest iteration pace. The M2.5 model has already been deployed internally at MiniMax, where it autonomously completes 30% of overall tasks across functions including research and development, product management, sales, human resources, and finance. In coding scenarios specifically, M2.5-generated code now accounts for 80% of new code submissions.

The advancement addresses a critical bottleneck in enterprise AI adoption: the prohibitive cost of running sophisticated AI agents for extended periods. At 50 tokens per second output speed, the model costs just $0.30 per hour of continuous operation—pricing that MiniMax positions as one-tenth to one-twentieth the cost of competing models including Claude Opus, Gemini 3 Pro, and GPT-5.

The model's efficiency gains stem from optimized task decomposition capabilities and reduced token consumption during reasoning processes, enabling 37% faster task completion compared to its predecessor M2.1 on the SWE-Bench Verified benchmark.

Programming Capabilities Reach Industry Peak

M2.5 has achieved benchmark-leading results in core programming evaluations, matching performance levels of Claude Opus series models. The system scored 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench—the highest in the industry for multilingual tasks—and demonstrated comparable performance to Opus 4.5 on the upgraded VIBE Pro benchmark.

The model exhibits what MiniMax characterizes as "architect-level thinking," demonstrating native specification behavior by proactively decomposing functionality, structure, and UI design before writing code. This represents a shift from simple bug-fixing toward comprehensive system architecture planning.

Training encompassed over 10 programming languages including Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart, and Ruby, across hundreds of thousands of real-world environments. The model handles full development lifecycles from zero-to-one system design and environment setup through system development, feature iteration, code review, and system testing across Web, Android, iOS, Windows, and Mac platforms.

Cross-scaffold generalization testing revealed strong performance consistency. On the Droid framework, M2.5 achieved 79.7% pass rate, exceeding M2.1's 71.3% and Opus 4.6's 78.9%. On OpenCode, it scored 76.1%, surpassing M2.1's 72.0% and Opus 4.6's 75.9%.

Enhanced Search and Tool Integration

M2.5 reached top-tier industry performance on BrowseComp (76.3%) and Wide Search benchmarks while demonstrating improved generalization across unfamiliar scaffold environments. The company developed RISE (Realistic Interactive Search Evaluation) to measure model performance on authentic expert-level search tasks, where M2.5 showed what the company describes as "excellent" performance.

Decision-making efficiency improved notably over previous generations. M2.5 achieved superior results using approximately 20% fewer search iterations across BrowseComp, Wide Search, and RISE tasks compared to M2.1, indicating more streamlined problem-solving pathways rather than brute-force approaches.

Professional-Grade Office Productivity

MiniMax collaborated with senior practitioners in finance, law, and social sciences to develop capabilities for delivering production-ready outputs in office scenarios. These domain experts contributed requirements, feedback, standards definition, and data construction, embedding industry tacit knowledge into the training process.

The model demonstrated significant capability improvements in advanced office scenarios including Word document processing, PowerPoint presentation creation, and Excel financial modeling. Using the internal Cowork Agent evaluation framework (GDPval-MM), which assesses output quality and process professionalism through pairwise comparison while monitoring token costs, M2.5 achieved a 59.0% average win rate against mainstream competing models.

Speed and Cost Economics Transform Agent Viability

M2.5's efficiency derives from three integrated factors: the model's native 100 tokens-per-second inference speed—nearly double current mainstream models—optimized complex task decomposition capabilities, and reduced token consumption during reasoning processes developed through reinforcement learning.

On SWE-Bench Verified tasks, M2.5 consumed an average 3.52 million tokens per task compared to M2.1's 3.72 million tokens. End-to-end execution time decreased from 31.3 minutes to 22.8 minutes, matching Claude Opus 4.6's 22.9-minute average.

The model offers two versions with identical capabilities but different speed-price configurations. The fast version operating at approximately 100 tokens per second charges 0.30permillioninputtokensand0.30permillioninputtokensand2.40 per million output tokens. The 50-tokens-per-second version halves the output pricing. At 100 tokens per second, one hour of continuous operation costs 1;at50tokenspersecond,1;at50tokenspersecond,0.30. This pricing enables four agents to run continuously for one year at $10,000 total cost.

Fastest Industry Progress Through Reinforcement Learning

Over the past 108 days, MiniMax released M2, M2.1, and M2.5 in succession, maintaining what the company characterizes as the industry's fastest improvement rate on SWE-Bench Verified compared to Claude, GPT, and Gemini model families.

The company attributes this acceleration to large-scale reinforcement learning, which significantly enhanced model capabilities and generalization across scaffolds and environments. Through co-design of its Agent RL framework, algorithms, reward mechanisms, and engineering optimizations, MiniMax achieved efficient optimization across arbitrary agent scaffolds and environments, training on hundreds of thousands of scaffolds including numerous internal real-world tasks. The company reports near-linear capability scaling with increased compute and task volume.

The proprietary Forge framework, designed as a native Agent RL system, introduces an intermediate layer that completely decouples underlying training and inference engines from agents, supporting integration of arbitrary agents and enabling optimization of model generalization across agent scaffolds and tools. Optimized asynchronous scheduling strategies balance system throughput with sample off-policyness, while tree-structured training sample merging achieved approximately 40x training acceleration.

Algorithmically, MiniMax employs its CISPO algorithm proposed earlier in 2026 to ensure Mixture-of-Experts model stability during large-scale training. To address credit assignment challenges in long-context agent scenarios, the company introduced process reward mechanisms for full-chain quality monitoring. The system directly estimates task completion time in real environments as a reward signal, balancing model effectiveness with response speed.

Deployment and Availability

M2.5 has been fully deployed across MiniMax Agent, integrating standardized Office Skills for Word formatting, PowerPoint editing, and Excel calculations. In MAX mode, the agent automatically loads corresponding Office Skills based on file type. Users can combine Office Skills with domain-specific expertise to create reusable experts for particular task scenarios.

Users have created over 10,000 custom experts on the MiniMax Agent platform, with continued rapid growth. MiniMax has developed multiple deeply optimized, ready-to-use expert suites for high-frequency scenarios including office work, finance, and programming.

The M2.5 model is now available across all MiniMax products including MiniMax Agent, M2.5 API access, and Coding Plan subscriptions. Model weights will be open-sourced on HuggingFace to support local deployment.