Alibaba Qwen3.7-Max has launched automatic implicit caching, reducing input costs by up to 80% According to Beating monitoring, the Alibaba Qwen team announced that automatic implicit caching will be enabled by default for its flagship model Qwen 3.7-Max on the Alibaba Cloud Bailian platform. Developers can directly enjoy cache cost reduction without modifying code or specifying additional parameters. Under the new billing mechanism, the system will automatically identify and extract duplicate context prefixes from requests. Once a cache hit occurs, the input token fee for the hit portion will only be charged at 20% of the original unit price, directly waiving 80% of the input cost. Implicit caching directly targets the huge overhead in long text and agent scenarios. Qwen3.7-Max, which has a long context window of 1 million tokens, requires frequent and repetitive reading of large code bases or knowledge documents when running high-level tasks such as autoencoder. After testing Qwen3.7, a developer reported that in less than an hour, they spent nearly 1 million tokens building a demo of the Tank Wars webpage. If we let the intelligent agent autonomously perform code review and loop iteration in the background, the daily usage can easily surge to hundreds of millions of tokens. The internal competition among peers in cache pricing is another direct incentive for Alibaba to lower prices. Previously, DeepSeek V4 Pro attracted a large number of developers with its extremely low cache hit price. After announcing a permanent price reduction at the end of May, DeepSeek V4 Pro's cache hit billing was reduced to only $0.003625 per million tokens (approximately RMB 0.025), which is equivalent to directly waiving 99.17% of the cost based on the standard input price. A large number of developers have collaborated with exclusive tools such as Reasonix to push the cache hit rate of a single session to the limit of 99%, resulting in almost zero running bills for long session agents. Faced with competitive pressure, Qwen3.7-Max not only launched implicit caching without any configuration, but also retained the explicit caching mode that requires manual declaration of the cache_comtrol identifier. Compared to automatic caching, explicit caching has higher hit certainty, with a hit cost as low as 10% of the standard input unit price (50% off). However, when creating a cache for the first time, a premium of 125% is required, and the cache block has a lifespan of only 5 minutes (which can be re timed during each launch). [Original link]