Qwen3.8 Max Prime

Qwen3.8 Max Prime is Alibaba's own version of the speed-tier pattern Z.ai has been running on GLM-5.3: the same model, a faster lane, and a price that maps to the multiple. Released September 23, 2026, it serves identical weights to Qwen3.8-Max, the 2.4 trillion parameter mixture-of-experts activating roughly 95 billion per token, with the same 1 million token context window and native text, image, and video input. Alibaba lists it at $4 per million input tokens and $12 output, exactly double the $2 and $6 the standard endpoint charges, and its own documentation is explicit that model-supported capabilities and usage restrictions are the same as the original model. What the premium buys is throughput: Alibaba claims output-speed-sensitive scenarios see TPS raised to 1.5 to 2 times the standard API, with a soft ceiling rather than a hard rate limit, so a request is not throttled as long as the platform has spare resources. No controlled independent measurement of that speed claim has been published yet, and early OpenRouter traffic figures cited by third parties showed a far smaller gap (about 40 versus 37 tokens per second), so treat it as a vendor number until someone benchmarks the two lanes side by side. Every benchmark score published for Qwen3.8-Max, including the 86.6 on Terminal-Bench 2.1 and the split software-engineering picture against Fable 5 and GPT-5.6 Sol, applies unchanged to Prime, because nothing about the model itself moved.