Qwen 3.8 27B API Pricing & Cost Calculator

How this is calculated

Qwen 3.8 27B is priced at $0.575 per million input tokens and $3.45 per million output tokens on Alibaba Cloud International, the proprietary hoster, with an 80% prompt caching discount ($0.115 per million cached reads) and a 50% batch discount ($0.2875/$1.725). Third-party hosts on OpenRouter undercut that to about $0.45/$3.20, but with weaker latency and uptime - Alibaba's own endpoint runs about 41 tokens/sec at 99.92% uptime, roughly double the throughput of the cheaper resellers. As with the rest of Alibaba's catalog, cache and batch discounts cannot apply simultaneously.

Verdict

At $0.575/M input and $3.45/M output a 100K + 5K call costs about $0.075, under a third of Qwen 3.8 Max's $0.23 and roughly a tenth of GPT-5.6 Sol's $0.65. Cached input drops to $0.115/M, which drives stable-prompt utility workloads (classification, routing, formatting) almost to zero on the input side. Cheaper third-party OpenRouter hosts list the same weights around $0.45/$3.20, but they post 15-24 tokens/sec and 97-99% uptime against Alibaba's 41 tokens/sec and 99.92%, so the 20-28% price cut buys a measurable reliability drop. Self-hosting is the other alternative: the weights run on a 48 GB GPU at about 45 GB of VRAM for 128K context.