MiniMax M2.5 has credible coding-oriented positioning, tool use, hosted API support, and open-weight deployment options in official MiniMax materials. Its reported 80.2% SWE-bench Verified result supports a coding score above GLM-4.6 (79), although it is vendor-reported, used internal infrastructure and Claude Code, and is not directly comparable to independently standardized runs. It therefore remains below DeepSeek-V3.2 (84), Kimi K2.5 (86), and frontier OpenAI/Anthropic anchors (95–96) for coding and agents. Artificial Analysis reports a 34 Intelligence Index, 75.3 output tokens/s, 1.70s TTFT, and $0.30/$1.20 per million input/output tokens. Those figures support strong value and above-median speed relative to the calibration set, though not DeepSeek-V3.2’s cost leadership. The supplied evidence establishes a text-focused model rather than broad native multimodal capability, so multimodal/I-O is held at the text-model floor. Open weights and API documentation improve developer experience, but supplied sources do not establish license terms or mature ecosystem depth. Independent coverage is incomplete: M2.5 is absent from the checked DeepSWE and LiveCodeBench leaderboards, and no Terminal-Bench or Aider result is supplied. On Arena Hard Prompts it ranks 143rd at 1415±5 from 25,350 votes, indicating meaningful exposure but modest preference relative to leading models. This evidence profile warrants conservative reasoning, adoption, and evidence-quality scores despite favorable vendor claims.
