MiniMax releases M2.5 coding and agent model
MiniMax says its new M2.5 model matches Claude Opus 4.6 on agentic coding speed and some benchmarks at a fraction of the cost, though all of the figures come from the company itself.
Chinese AI lab MiniMax has announced MiniMax-M2.5, the latest model in its M2 series. The company says the model was trained with reinforcement learning across hundreds of thousands of complex real-world environments.
MiniMax reports scores of 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench and 76.3% on BrowseComp (with context management). To test how well the model handles unfamiliar agent harnesses, MiniMax also ran SWE-Bench Verified through third-party ones. It scored 79.7 on Droid against 78.9 for Claude Opus 4.6, and 76.1 on OpenCode against Opus 4.6's 75.9.
Speed and price are the main pitch. MiniMax says M2.5 finished the SWE-Bench Verified evaluation 37% faster than M2.1, matching Opus 4.6. It used an average of 3.52 million tokens per task, compared with 3.72 million for M2.1. The model is served natively at 100 tokens per second, which the company says is nearly twice the rate of other frontier models. MiniMax puts the cost of running it continuously for an hour at $1 at that speed, or $0.30 at 50 tokens per second.
The company says coding training covered more than 10 languages in over 200,000 environments. On agentic search tasks, it reports using about 20% fewer rounds than M2.1. For office work, it says it worked with finance, law and social-science professionals and reports a 59.0% average win rate on its internal GDPval-MM evaluation.
Several of these results come from MiniMax's own benchmarks, including VIBE-Pro, RISE and GDPval-MM. None of the figures have been independently verified. The announcement excerpt does not give details on model size, licensing or availability.