Qwen3.8-Max vs Alibaba international
Qwen3.8-Max vs Alibaba international
Alibaba’s Singapore (international) list for Qwen3.8-Max is about $2.00 input / $0.25 cached input / $6.00 output per 1M tokens. That is a solid flagship price. On a month that sends every token to Max, switching hosts for a few cents per million barely moves the bill. What does move it is putting the easy steps on Flash and keeping Max for the hard ones, on one OpenAI-compatible key.
阿里巴巴新加坡(国际)站 Qwen3.8-Max 的定价约为每百万 token 输入 2.00 美元、缓存输入 0.25 美元、输出 6.00 美元。这是一个稳健的旗舰级定价。在一个所有 token 都发送给 Max 的月份里,为了每百万 token 几美分的差价而更换主机商几乎不会改变账单总额。真正能改变账单的是将简单步骤交给 Flash 处理,并将 Max 保留给困难任务,且仅使用一个兼容 OpenAI 的密钥。
List prices used (USD per 1M tokens, checked 2026-11-06)
| Model | Source | Input | Cached input | Output |
|---|---|---|---|---|
| Qwen3.8-Max | Alibaba Model Studio Singapore | ~$2.00 | ~$0.25 | ~$6.00 |
| Qwen3.8-Max | IteraCompute | $1.95 | $0.25 | $5.70 |
| Qwen3.8-Flash | Alibaba Model Studio Singapore | $0.15 | — | $0.47 |
| Qwen3.8-Flash | IteraCompute | $0.11 | $0.016 | $0.38 |
所用标价(美元/百万 token,核查日期 2026-11-06)
| 模型 | 来源 | 输入 | 缓存输入 | 输出 |
|---|---|---|---|---|
| Qwen3.8-Max | 阿里巴巴 Model Studio 新加坡 | ~$2.00 | ~$0.25 | ~$6.00 |
| Qwen3.8-Max | IteraCompute | $1.95 | $0.25 | $5.70 |
| Qwen3.8-Flash | 阿里巴巴 Model Studio 新加坡 | $0.15 | — | $0.47 |
| Qwen3.8-Flash | IteraCompute | $0.11 | $0.016 | $0.38 |
Alibaba Flash cache rates depend on implicit vs explicit mode on Model Studio; the tables below only apply a Flash cache price on the IteraCompute rows. Max cache is shown at $0.25 on both sides so the host comparison is apples-to-apples. Pull the live Model Studio sheet before you budget.
阿里巴巴 Flash 的缓存费率取决于 Model Studio 中的隐式或显式模式;下表仅在 IteraCompute 行中应用了 Flash 缓存价格。Max 缓存价格在双方均显示为 0.25 美元,以便进行同类比较。在制定预算前,请获取 Model Studio 的实时价格表。
One month: 100M input + 10M output, all on Max
| Cache-hit rate | Alibaba Singapore Max | IteraCompute Max |
|---|---|---|
| 0% | $260.00 | $252.00 |
| 50% | $172.50 | $167.00 |
| 80% | $120.00 | $116.00 |
| 90% | $102.50 | $99.00 |
| Host choice on Max alone saves about 3% here. Useful, not decisive. |
一个月:1 亿输入 + 1000 万输出,全部使用 Max
| 缓存命中率 | 阿里巴巴新加坡 Max | IteraCompute Max |
|---|---|---|
| 0% | $260.00 | $252.00 |
| 50% | $172.50 | $167.00 |
| 80% | $120.00 | $116.00 |
| 90% | $102.50 | $99.00 |
| 仅在 Max 上选择主机商可节省约 3% 的费用。有用,但非决定性因素。 |
Same month, 70% Flash / 30% Max (easy steps vs hard steps) Workload assumption: 70% of tokens on Flash, 30% on Max, same 10:1 input-to-output ratio on both. That is a typical agent split (classify / summarise / extract / format on Flash; plan / patch / final answer on Max).
同月,70% Flash / 30% Max(简单步骤 vs 困难步骤) 工作负载假设:70% 的 token 使用 Flash,30% 使用 Max,两者输入输出比均为 10:1。这是典型的智能体分工(在 Flash 上进行分类/总结/提取/格式化;在 Max 上进行规划/补丁/最终回答)。
| Setup | 0% cache | 50% | 80% | 90% |
|---|---|---|---|---|
| 70/30 on Alibaba Singapore (Flash cache not applied) | $91.79 | — | — | — |
| 70/30 on IteraCompute (one key) | $85.96 | $57.17 | $39.90 | $34.14 |
| All Max on Alibaba Singapore | $260.00 | $172.50 | $120.00 | $102.50 |
| 设置 | 0% 缓存 | 50% | 80% | 90% |
|---|---|---|---|---|
| 阿里巴巴新加坡 70/30(未应用 Flash 缓存) | $91.79 | — | — | — |
| IteraCompute 70/30(一个密钥) | $85.96 | $57.17 | $39.90 | $34.14 |
| 阿里巴巴新加坡全 Max | $260.00 | $172.50 | $120.00 | $102.50 |
The number to notice: routing 70% of tokens to Flash cuts the all-Max Alibaba bill by about two thirds even before you change hosts ($260 → ~$92 at 0% cache). On one OpenAI-compatible key the same mix is a single model= flip per step — no second client, no second top-up.
值得注意的数据是:将 70% 的 token 路由到 Flash,即使在更换主机商之前,也能将全 Max 的阿里巴巴账单削减约三分之二(在 0% 缓存时从 260 美元降至约 92 美元)。在使用一个兼容 OpenAI 的密钥时,同样的组合只需在每一步切换 model 参数即可——无需第二个客户端,也无需第二次充值。
How to route without a fancy router
from openai import OpenAI
client = OpenAI(base_url="https://api.iteracompute.com/v1", api_key="YOUR_KEY")
EASY = {"classify", "summarize_tool_output", "extract", "commit_message", "format_check"}
def run(step: str, messages: list[dict]) -> str:
model = "qwen/qwen3.8-flash" if step in EASY else "qwen/qwen3.8-max"
r = client.chat.completions.create(model=model, messages=messages)
return r.choices[0].message.content
如何无需复杂的路由器进行路由 (代码见上文)
The same pattern works with GLM-5.3-Flash or DeepSeek V4.1 Flash as the cheap side if that family fits the easy steps better. Keep a fallback: if a Flash step fails validation, retry once on Max and count those retries in the bill.
如果 GLM-5.3-Flash 或 DeepSeek V4.1 Flash 更适合简单步骤,同样的模式也适用于它们作为廉价端。请保留一个回退机制:如果 Flash 步骤验证失败,在 Max 上重试一次,并将这些重试计入账单。
Caveats Thinking / reasoning tokens usually bill as output. Default high effort on Max inflates the output column faster than the input price suggests. Put the stable prefix (system prompt, tools, repo context) first so cache hits stay high. A request ID at the top of the system prompt resets the hit rate. This is not a claim that any host is the cheapest in the world. fp8 / promo / multi-provider pin setups can undercut a single-host mix. Measure on your own traffic.
注意事项 思考/推理 token 通常按输出计费。Max 默认的高努力程度会导致输出列的增长速度快于输入价格所暗示的水平。将稳定的前缀(系统提示词、工具、仓库上下文)放在最前面,以保持较高的缓存命中率。系统提示词顶部的请求 ID 会重置命中率。这并非声称任何主机商是全球最便宜的。fp8 / 促销 / 多供应商固定设置可能会比单一主机组合更便宜。请根据您自己的流量进行衡量。
This is IteraCompute. We serve qwen/qwen3.8-max and qwen/qwen3.8-flash through one OpenAI-compatible API. On 2026-11-06 our Max price is $1.95 input / $0.25 cached input / $5.70 output per 1M tokens, a little under Alibaba’s international list (~$2.00 / $0.25 / $6.00). Flash is $0.11 / $0.016 / $0.38. The combined bill comes from routing, not from shaving Max by a few cents: Iteracompute Qwen3.8 Max
这是 IteraCompute。我们通过一个兼容 OpenAI 的 API 提供 qwen/qwen3.8-max 和 qwen/qwen3.8-flash 服务。在 2026 年 11 月 6 日,我们的 Max 定价为每百万 token 输入 1.95 美元 / 缓存输入 0.25 美元 / 输出 5.70 美元,略低于阿里巴巴的国际定价(约 2.00 美元 / 0.25 美元 / 6.00 美元)。Flash 定价为 0.11 美元 / 0.016 美元 / 0.38 美元。综合账单的节省来自于路由策略,而不是通过在 Max 上节省几美分:Iteracompute Qwen3.8 Max。
Sources: Alibaba Cloud Model Studio Singapore pricing (Qwen3.8-Max / Qwen3.8-Flash), IteraCompute /v1/models (fetched 2026-11-06, cache bypassed).
来源:阿里云 Model Studio 新加坡定价(Qwen3.8-Max / Qwen3.8-Flash),IteraCompute /v1/models(获取于 2026-11-06,已绕过缓存)。