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.
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 |
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.
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.
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).
| 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 |
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.
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
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.
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.
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
Sources: Alibaba Cloud Model Studio Singapore pricing (Qwen3.8-Max / Qwen3.8-Flash), IteraCompute /v1/models (fetched 2026-11-06, cache bypassed).
Top comments (0)