Alibaba's new 2.4 trillion parameter flagship advertises a 1,000,000 token context window, but its documented maximum input length is 991,808 tokens, and 983,616 in thinking mode. Teams sizing long context jobs against the round number will hit a wall. The bigger lever is the cache: implicit cached input is priced at a quarter of standard input.
Qwen3.8-Max landed on 3 August as Alibaba's largest model, a Mixture of Experts flagship with 2.4 trillion total parameters. The specification everyone repeated was the context window: 1,000,000 tokens. That number is real, and it is also not the number you can actually fill with your own content.
Alibaba's own model documentation separates two figures that most coverage collapsed into one. The context window is 1,000,000 tokens. The maximum input length is 991,808 tokens. The difference is a little over 8,000 tokens, and it exists because the context window has to hold the response as well as the prompt. Turn on thinking mode and the input ceiling drops further, to 983,616 tokens, because the reasoning trace has to live somewhere too.
This matters more than the size of the gap suggests, because of how long context jobs fail. A pipeline that chunks a corpus to a round 1,000,000 tokens does not degrade gracefully when it goes over. It errors, and it errors after you have already paid to assemble and transmit the payload. Teams that size to the advertised window and discover the real ceiling in production are debugging a rejected request rather than a slow one, and the failure lands on the largest and most expensive calls in the workload.
The economics are where the real decision sits, and they run in the opposite direction from where most teams look. In the Singapore region Alibaba lists standard input at 2 dollars per million tokens and output at 6 dollars per million. Implicit cached input is listed at 0.25 dollars per million tokens. That is one eighth the standard input price, and on a near million token prompt the difference between a cache hit and a cache miss is most of the bill.
That ratio inverts the usual optimization instinct. The reflex on a long context workload is to shrink the prompt. The higher leverage move is to make the large prompt stable, so the unchanging bulk of it stays cacheable across calls while only the variable tail changes. A 900,000 token document that is resent identically on every query and a 900,000 token document that gets reshuffled each time cost dramatically different amounts to serve, and the difference is a property of how you assemble the prompt rather than of the model.
Try it today
- Set your hard input ceiling to 991,808 tokens, not 1,000,000, and to 983,616 if you are running with thinking mode enabled.
- Measure your actual token count before dispatch rather than estimating from character or word counts, and fail the request locally instead of paying to have the API reject it.
- Restructure long prompts so the stable bulk comes first and the variable portion comes last, which is what makes the fixed segment eligible for cache reuse across calls.
- Price your workload both ways before committing: standard input at 2 dollars per million against implicit cached input at 0.25 dollars per million in the Singapore region.
- Reserve the top of the window for headroom. Running a long context job at 99 percent of the documented input ceiling leaves nothing for prompt drift as your template evolves.