Qwen3.7-Max's Automatic Caching Cuts Costs by 80%

Alibaba's Qwen team has announced that its flagship model Qwen3.7-Max now automatically enables implicit caching on the Alibaba Cloud Bailian platform. Developers get cost savings without modifying code or specifying parameters.

Under the new pricing, the system automatically identifies and extracts repeated context prefixes from requests. When a cache hit occurs, the hit portion costs only 20% of the original per-token price — an 80% savings.

The implicit cache targets the massive costs of long-context and agent scenarios. Qwen3.7-Max supports a 1-million-token context window, and during autonomous coding tasks, it frequently re-reads large codebases or knowledge documents. One developer testing Qwen3.7 reported consuming nearly 1 million tokens in under an hour building a Tank Battle web demo. If an agent runs background code reviews and iterative loops unattended, daily usage can easily hit hundreds of millions of tokens.

Competitive pressure from rivals on cache pricing is another direct factor driving Alibaba's price cut. Earlier, DeepSeek V4-Pro attracted developers with its extremely low cache hit price. After a permanent price cut announced at the end of May, DeepSeek V4-Pro's cache hit fee dropped to just $0.003625 per million tokens (about 0.025 RMB), effectively eliminating 99.17% of standard input costs. Many developers, using specialized tools like Reasonix, have pushed cache hit rates as high as 99% in single sessions, bringing the cost of long-running agent chats close to zero.

In response to competition, Qwen3.7-Max not only launched zero-configuration implicit caching but also retains an explicit caching mode that requires manually declaring a cache_control flag. Compared to automatic caching, explicit caching offers higher hit certainty, with hit costs as low as 10% of the standard input price. However, creating a cache for the first time incurs a 125% premium, and cache blocks have a lifespan of just five minutes (refreshed on each hit).