AI API Price War: Low-Cost Models, Search APIs and Multi-Model Routing
Key Questions
What are the new pricing details for DeepSeek V4?
DeepSeek V4 Pro/Flash offers a permanent 75% price cut at $0.435 per million input tokens and $0.0036 per million cache tokens. It features a 1.6T/862B MIT MoE model with 1M context length and SOTA coding performance.
How have MiMo models changed in price and performance?
MiMo V2.5 API received permanent price cuts up to 99% at $0.0028 per million cache hits, leading to 5-8x usage growth. The new MiMo-V2.5-Pro-UltraSpeed reaches over 1000 tokens per second on a 1T model but requires application-based access.
Which new models have been open-sourced recently?
MiMo Code is now open-source for low-cost self-hosted coding agents, while GLM-5.2 open-weight tops Terminal-Bench at over 80% and ranks second globally on Code Arena. Kimi K2.7-Code is also open-source and claimed to be 100x cheaper than Fable 5.
What is OpenRouter Fusion and its impact?
OpenRouter Fusion API enables multi-model blending for frontier-level performance at half the cost. OpenRouter itself reached unicorn status with a $1.3B valuation.
How does Grok Imagine 1.5 compare in pricing to competitors?
Grok Imagine 1.5 is priced at $4.20 per minute for 720p video, making it 86% cheaper than Sora 2 Pro.
GPT-6.1 Terra 报道 API 定价为输入 $1、输出 $5/百万 token，并有 Batch/Flex 折扣；DeepSeek Search/Fetch、Exa 3.0 Lite 和 Qwen3.8 Flash 等报价仍需以账户实价和真实工作流核验。本轮关于 DeepSeek agent 的报道是研究原型，未带来新的 API 定价或生产性能证据。