GLM-4 和 MiniMax M2.5 定价 : API成本、模型与对比
On this page
Two of the most cost-effective AI model families right now come from Chinese labs: Zhipu AI’s GLM series and MiniMax’s M2.5. Both offer frontier-level performance at a fraction of what you would pay for Claude or GPT. Here is what they actually cost as of February 2026.
If you are comparing AI model costs more broadly, our Claude AI定价指南 covers Anthropic’s full lineup including the new Opus 4.6.
GLM-4 系列定价
智谱AI（GLM模型背后的公司）通过其Z.AI平台运营。“GLM-4”家族已显著扩展，最新模型包括GLM-5、GLM-4.7以及多个GLM-4.5变体。
以下所有价格均以每100万个令牌计价，单位为美元。
| 模型 | 输入 | 缓存输入 | 输出 | 上下文窗口 |
|---|---|---|---|---|
| GLM-5 | $1.00 | $0.20 | $3.20 | 128K |
| GLM-5-Code | $1.20 | $0.30 | $5.00 | 128K |
| GLM-4.7 | $0.60 | $0.11 | $2.20 | 200K |
| GLM-4.5 | $0.60 | $0.11 | $2.20 | 128K |
| GLM-4.5-X | $2.20 | $0.45 | $8.90 | 128K |
| GLM-4.5-Air | $0.20 | $0.03 | $1.10 | 128K |
| GLM-4.7-Flash | 免费 | 免费 | 免费 | 128K |
| GLM-4.5-Flash | 免费 | 免费 | 免费 | 128K |
这里引人注目的是：GLM-4.7-Flash和GLM-4.5-Flash是免费的。没有陷阱，没有每日速率限制配额。智谱提供这些作为亏本销售产品以促进采用。
GLM-4.7是旗舰开源模型。它在SWE-bench Verified上得分73.8%，在LiveCodeBench V6上得分84.9，使其在编码基准测试中领先于Claude Sonnet 4.5。200K的上下文窗口和128K的输出容量使其与成本高出5到10倍的模型具有竞争力。
缓存输入定价值得关注。如果您正在构建重复发送相同系统提示或参考文档的应用程序，GLM-4.7的缓存输入价格为$0.11/MTok，比标准输入价格便宜约80%。
MiniMax M2.5 定价
MiniMax于2026年2月12日发布了M2.5，它立即引起了轰动。该模型有两个版本：标准版和闪电版。
| 模型 | 输入 | 输出 | 上下文窗口 | 速度 |
|---|---|---|---|---|
| M2.5 Standard | $0.30/MTok | $1.20/MTok | 1M tokens | 50 tok/s |
| M2.5 Lightning | $0.30/MTok | $2.40/MTok | 200K tokens | 100 tok/s |
每百万令牌输入$0.30 / 输出$1.20，M2.5标准版的价格仅为Claude Opus 4.6（$5/$25）的1/63。这不是打字错误。
截至2026年2月中旬，M2.5在SWE-Bench Verified上得分80.2%，使其在公开基准测试中位居前列。它还在Multi-SWE-Bench上达到51.3%，在BrowseComp上通过上下文管理达到76.3%。
闪电版将输出价格翻倍，但也将吞吐量翻倍至每秒100个令牌。MiniMax声称您可以在100 tok/s的速度下连续运行模型一小时，费用约为$1。
M2.5是开源的，并在Hugging Face上可用，因此如果您有硬件，也可以自行托管。
与Claude和GPT的比较
这些模型与西方主要供应商相比如何？以下是价格和规格的并排比较。
| 模型 | 输入/百万令牌 | 输出/百万令牌 | 上下文 | SWE-bench |
|---|---|---|---|---|
| MiniMax M2.5 | $0.30 | $1.20 | 1M | 80.2% |
| GLM-4.7 | $0.60 | $2.20 | 200K | 73.8% |
| GLM-5 | $1.00 | $3.20 | 128K | N/A |
| Claude Sonnet 4.5 | $3.00 | $15.00 | 200K | ~70% |
| Claude Opus 4.6 | $5.00 | $25.00 | 200K | ~79% |
| GPT-4o | $2.50 | $10.00 | 128K | ~65% |
价格差距惊人。对于相同的$100预算：
- MiniMax M2.5：约8300万输出令牌
- GLM-4.7：约4500万输出令牌
- Claude Sonnet 4.5：约670万输出令牌
- Claude Opus 4.6：约400万输出令牌
在可比较的基准性能下，M2.5相对于Opus 4.6具有20倍的成本优势。
您应该选择哪个模型？
**如果您需要免费层级进行原型设计或低用量使用，请选择GLM-4.7-Flash。**它确实是免费的，没有隐藏配额，这使其非常适合在决定使用付费模型之前测试想法。
如果您需要一个价格适中且功能强大的通用模型，请选择GLM-4.7或GLM-5。$0.60/$2.20的定价远低于西方替代品，200K的上下文窗口可以处理长文档。
**如果成本是您的首要考虑因素，并且您需要前沿性能，请选择MiniMax M2.5标准版。**在$0.30/$1.20的价格下，它是其性能层级中最便宜的模型，优势显著。1M的令牌上下文窗口也是此价格点下最大的。
如果您需要实时应用程序的速度，请选择MiniMax M2.5闪电版。$0.30/$2.40的价格下，100 tok/s的吞吐量对于交互式编码代理和实时聊天应用程序具有竞争力。
**如果您需要最大程度的可靠性、完善的生态系统支持，或特定功能（如Claude的计算机使用或GPT的图像生成），请坚持使用Claude或GPT。**高昂的定价伴随着成熟的工具、详尽的文档和经过验证的生产稳定性。
Things to watch out for
Rate limits vary. Both Zhipu and MiniMax may apply different rate limits than what you get with OpenAI or Anthropic. Check the platform docs before planning high-volume deployments.
Latency from outside Asia. Both services run primarily from Chinese data centers. If your users are in North America or Europe, expect higher latency than with US-based providers. MiniMax offers some international endpoints, and GLM models are available through third-party providers like Fireworks and Novita.
Benchmark scores do not tell the whole story. M2.5’s 80.2% SWE-bench score is impressive, but real-world coding performance depends on your specific use case. Test with your actual workloads before committing.
Vision and multimodal costs differ. GLM-4.6V (the vision model) costs $0.30/$0.90 per MTok. MiniMax M2.5 currently supports text only. If you need image understanding, factor that into your comparison.
Using AI with your recordings
If you are processing meeting recordings or video content, per-token API costs can add up fast. A one-hour meeting transcript runs 10,000 to 15,000 tokens. Analyzing 100 meetings per month at Claude Opus rates ($5/$25 MTok) costs around $5 to $10 in API fees.
With ScreenApp, AI analysis is built into the platform. You get automatic transcription, AI summaries, and multi-language support without managing API keys or worrying about token budgets.
常见问题
How much does GLM-4 cost?
The GLM-4 family ranges from free (GLM-4.7-Flash, GLM-4.5-Flash) to $2.20 per million output tokens (GLM-4.7, GLM-4.5). The premium GLM-4.5-X model costs $8.90 per million output tokens. All pricing is through Zhipu AI’s Z.AI platform.
What is MiniMax M2.5 pricing?
MiniMax M2.5 Standard costs $0.30 per million input tokens and $1.20 per million output tokens. The Lightning version costs $0.30 input and $2.40 output per million tokens, with double the speed.
Is GLM-4 cheaper than ChatGPT?
Yes, significantly. GLM-4.7 costs $0.60/$2.20 per million tokens compared to GPT-4o at $2.50/$10.00. The Flash variants are free. Even GLM-5, the most expensive option, undercuts GPT-4o on both input and output pricing.
Is MiniMax M2.5 as good as Claude Opus?
On coding benchmarks, M2.5 scores 80.2% on SWE-Bench Verified versus Opus 4.6’s approximately 79%. Performance on other tasks varies. M2.5 costs roughly 1/20th of Opus 4.6, making it worth testing for your specific use case.
Can I self-host GLM-4 or MiniMax M2.5?
Both model families are open-source. GLM-4.7 is available on Hugging Face from Zhipu AI (zai-org), and MiniMax M2.5 is available from MiniMaxAI on Hugging Face. Self-hosting requires significant GPU resources but eliminates per-token costs entirely.
What context window does MiniMax M2.5 support?
M2.5 Standard supports up to 1 million tokens of context. The Lightning version supports 200K tokens. Both can output up to 131K tokens per response.
