MiniMax M2.5
MiniMax M2.5 via the Vercel AI Gateway: state-of-the-art coding, agentic tool use, search, and office work. 228B MoE, 196k context. Function calling supported. Served only by zero-data-retention providers. Does not support web search or image input.
You always get the exact model you pick — we never silently route you to another.
About MiniMax M2.5
When Claude Opus 4.6 costs roughly 13 times more per token and MiniMax M2.5 matches its SWE-Bench Verified score (80.2% vs. 80.8%), the economics of coding-heavy workloads shift considerably. Built on a 230-billion-parameter Mixture of Experts architecture with only 10 billion parameters active per inference, M2.5 reaches frontier-level software engineering performance without frontier-level pricing — at $0.15 per million input tokens for the standard variant. People who run it in agent harnesses and automated coding pipelines appreciate both the cost curve and the model's tendency to approach problems like a software architect: decomposing features, designing structure, and planning before writing. Its 76.8% score on multi-turn tool-calling benchmarks and strong web navigation results make it a practical choice for autonomous workflows. The honest caveat: pure reasoning and general knowledge tasks are noticeably weaker than the coding benchmarks suggest, and some users report the model getting stuck in agent loops or ignoring specific formatting constraints. For bulk coding automation and office document generation, though, the price-to-capability ratio is hard to match.
Best for
- High-volume software engineering and code generation across Web, Android, iOS, and Windows platforms
- Autonomous agent workflows requiring multi-turn tool calling and web navigation
- Office automation including Word, PowerPoint, and Excel document generation and manipulation
- Cost-sensitive production deployments where frontier-level coding quality is needed at scale
- ML research pipelines and complex multi-step workflow automation
Specifications
| Provider | MiniMax |
|---|---|
| Released | 2026-02 |
| Intelligence | Medium |
| Speed | Fast |
| Context window | 196,600 tokens |
| Max output | 196,608 tokens |
| Knowledge cutoff | Not specified |
| Input price | $0.30 / 1M tokens |
| Output price | $1.20 / 1M tokens |
| Request cost | 2 base requests |
| Plan tier | Base |
| Input | Text |
| Output | Text |
| Features | Function calling supported, Served through the Vercel AI Gateway (minimax/minimax-m2.5), Zero data retention and no prompt training |
| Model ID | minimax-m2p5 |
Frequently asked questions
The standard M2.5 variant costs $0.15 per million input tokens and $1.15 per million output tokens. The M2.5-Lightning variant (100 tokens/second throughput) is $0.30 input / $2.40 output per million tokens. Both variants support prompt caching to reduce costs on repeated-context workloads.
204,800 tokens (roughly 307 A4 pages), with a maximum output of 196,608 tokens.
M2.5 scores 80.2% on SWE-Bench Verified versus Claude Opus 4.6's 80.8% — essentially equivalent — while costing approximately 13 times less per token, making it a strong alternative for teams running large-scale coding workloads.
General reasoning and broad knowledge benchmarks (MMLU: 85%) are notably weaker than its coding scores, as the model was deliberately optimized for software engineering and agent tasks. Users also report agent loop issues and difficulty following strict formatting constraints.
Yes — standard M2.5 runs at ~50 tokens/second at lower cost, while M2.5-Lightning runs at ~100 tokens/second for roughly double the price. Use Lightning for latency-sensitive or interactive applications; use standard M2.5 for batch processing where throughput flexibility is acceptable.
Yes. MiniMax released M2.5 as open weights under the MIT license, allowing local deployment and fine-tuning for commercial use.