MiniMax Releases M2.5 and M2.5 Lightning at Fraction of Frontier Model Cost
On February 12, 2026, Shanghai-based MiniMax released M2.5 and M2.5 Lightning. The models achieve near state-of-the-art results on coding, agentic tool use, and productivity benchmarks at roughly one-tenth to one-twentieth the cost of leading models like Claude Opus 4.6, with Lightning variant offering 100 tokens per second throughput.
TLDR
MiniMax launched M2.5 and M2.5 Lightning on February 12, 2026. The models post strong scores including 80.2% on SWE-Bench Verified, leading or near-leading on Multi-SWE-Bench and BrowseComp. M2.5 Lightning runs at 100 tokens/second (2x typical frontier throughput) at $0.30/M input and $2.40/M output tokens; the standard variant is even lower priced. Both support caching and are positioned for continuous agent operation at very low cost (around $1/hour in some estimates).
Performance and Pricing
The release emphasizes real-world productivity tasks: coding, agentic tool use, search, and office work. The company states the models were trained extensively with reinforcement learning across hundreds of thousands of complex environments. Weights are available on Hugging Face under a modified MIT license for the open variants, with full API access also offered.
M2.5 Lightning targets speed-critical deployments while maintaining identical capability to the standard release. Both versions are significantly cheaper than comparable US frontier models on output pricing, shifting the economics for always-on agents and high-volume enterprise workloads.
Why this story matters
Dramatic cost reductions at near-frontier performance levels accelerate the diffusion of capable AI beyond hyperscale labs and large enterprises. When a model can sustain high-quality agentic work for roughly $1 per hour of continuous operation, new classes of applications — persistent personal agents, always-on monitoring, mass parallel experimentation — become practical. The open-weights release also intensifies the global competition between proprietary high-margin offerings and lower-cost, widely distributable alternatives, with implications for adoption curves, safety surface area, and supply-chain geopolitics.
Sources
- MiniMax official announcement: “MiniMax M2.5: Built for Real-World Productivity,” dated February 12, 2026. https://www.minimax.io/news/minimax-m25 (benchmark numbers, pricing, speed claims, licensing).
- VentureBeat and independent coverage confirming cost comparisons (1/10–1/20 of Opus/Gemini/GPT equivalents) and benchmark proximity (Feb 12, 2026).
- Hugging Face model cards and community reports on M2.5 availability and throughput.
Featured Image Alt Text
MiniMax M2.5 model release graphic highlighting low cost, high speed, and agentic productivity benchmarks
Tags
MiniMax, M2.5, Low Cost AI, Efficiency, Agentic Models, Open Weights, China AI