A lower price makes Qwen4-Max worth testing, but replacing model is only part of this migration. Your old answer limit now includes reasoning, and effort values do not all keep their meaning. Start with those settings, then measure the cost of a result your application accepts.
The configuration difference
The Qwen4-Max documentation and Qwen3.8-Max model page list the same 1,000,000-token context and text/image/video inputs. Non-thinking output is 131,072 on both. That does not mean their thinking-mode budgets are interchangeable.
| Setting | Qwen3.8-Max | Qwen4-Max |
|---|---|---|
| Native effort | low, medium, xhigh | low, high, xhigh |
| Default effort | xhigh | high |
| max_tokens covers | Answer | Thinking + answer |
Qwen4-Max allows 262,144 tokens for thinking and answer combined. Qwen3.8-Max's 131,072 answer limit applies in thinking mode too, with a separate maximum chain-of-thought length of 262,144. Do not compare those different scopes as a simple multiplication of output capacity.
The Chat Completions reference recommends max_completion_tokens for new integrations. On Qwen4-Max it means the same as max_tokens; renaming the field without reconsidering the number changes nothing. Leave room for reasoning and the final answer, and check finish_reason. A length result means the budget was exhausted and may arrive before a useful answer.
Check effort mappings as well. Qwen4-Max maps medium to high; Qwen3.8-Max maps high to xhigh. When effort and thinking budget are both omitted, Qwen4-Max defaults to high and 131,072 thinking tokens. Never send reasoning_effort and thinking_budget together. An explicit budget must be positive; use the documented thinking switch to disable thinking. Keep historical reasoning in its own field, outside answer text.
Price the same usage first
At official Singapore USD rates, ordinary input falls from $2 to $1.50 per million tokens, and output including reasoning falls from $6 to $4.50. Each is 25% lower.
Across multiple requests, assume one million uncached input tokens and 100,000 output tokens including reasoning. Qwen3.8-Max costs $2 + $0.60 = $2.60; Qwen4-Max costs $1.50 + $0.45 = $1.95. These are cumulative totals, not one oversized request. The estimate excludes caches, tools, retries, and tax; other regions differ.
A task may consume different tokens after migration. Compare accepted results and total spend including retries before treating that 25% as an application saving.
Verify your route before switching traffic
The AIHubMix Qwen4-Max guide lists qwen4-max on Chat Completions but marks capabilities unverified. Vendor support for tools, structured output, Responses, or Messages does not prove your gateway path works.
Alibaba's Qwen4 explicit cache has automatic breakpoints and 5-minute or 1-hour TTL. Account for writes and reads separately. Optional thinking summaries require preserving encrypted history. Verify those controls on the route you intend to use; this article reports no new inference tests.
Choose one repository change, one long-document question, and a visual task if relevant. Fix prompts, tools, and acceptance criteria. Record correctness, total tokens, and failures; keep Qwen3.8-Max until the replacement passes. The AIHubMix comparison helps check listed specifications, while these tasks decide the upgrade.
Keep reading: the Qwen4-Max series
- For request-level migration steps: Qwen4-Max migration: fix the output budget before switching agents.
- For a second coding-model candidate: Qwen4-Max vs Kimi K3.1 Pro: compare cost per accepted patch.
Sources
- Alibaba Cloud: Qwen4-Max and Qwen3.8-Max.
- Alibaba Cloud: official pricing, Singapore USD rates.
- Alibaba Cloud: Chat Completions reference.
- Alibaba Cloud: Qwen4 cache and thinking summary.
- AIHubMix: Qwen4-Max guide and Qwen3.8-Max guide. Sources checked November 17, 2026, 02:33 UTC; no new inference tests.
Top comments (0)