Short answer: Qwen3.8 Max lists $2.00 per million input tokens and $6.00 per million output tokens on OpenRouter with a $0.25 cache-read rate and a one-million-token context. The batch variant is $1.15/$3.45 but caps at 262,144 tokens, Prime is $4.00/$12.00, and the Flash route is $0.15/$0.47. Price your real token mix on the route you can actually use rather than assuming the family's entry rate applies to your workload.
| Route | Prompt | Completion | Constraint |
|---|---|---|---|
| qwen/qwen3.8-max-09021,000,000 context | $2.00 | $6.00 | $0.25 cache read; the standard Max route. |
| qwen/qwen3.8-max-0902:batch262,144 context | $1.15 | $3.45 | Cheaper per token but a much smaller context window. |
| qwen/qwen3.8-max-prime1,000,000 context | $4.00 | $12.00 | Double the Max rate with a $0.50 cache read. |
| qwen/qwen3.8-flash1,000,000 context | $0.15 | $0.47 | High-volume route; verify quality on your own task set. |
| qwen/qwen3.8-27b1,048,575 context | $0.15 | $1.875 | Smaller model with a 12.5x output-to-input ratio. |
| qwen-api on together-aiserverless | $2.50 | $6.25 | APIDir accepted observation checked 2026-08-22. |
Four Qwen routes, four budgets
The Qwen3.8 catalog on OpenRouter contains routes that share a family name and nothing else about their economics. Max sits at $2.00 input and $6.00 output. Prime doubles both rates, which is the kind of premium a team pays for a specific capability, not a general default. Flash drops to $0.15/$0.47, and the 27B route prices output at $1.875 against $0.15 input. Each of those ratios implies a different workload fit.
Start by writing the ratio down. Max is 3:1 output-to-input, the 27B route is 12.5:1, and Prime is 3:1 at twice the absolute level. A document-summarization job that spends ten input tokens for every output token barely notices output pricing. A chat product with long answers does. The same route can look cheap or expensive depending on a ratio you can measure from your own usage logs before you commit.
Context limits decide the route before price does
The batch variant is priced at $1.15/$3.45, which is 42% below Max on both meters. It also caps the context at 262,144 tokens instead of one million. For a document pipeline that sends 300,000 tokens per request, the cheaper route simply cannot accept the workload; the choice is not a discount but a different product. This is the most common Qwen budgeting error: applying the batch rate to prompts that exceed the batch window.
Check the ceiling against your real distribution rather than your average. A pipeline averaging 180,000 tokens can still send a small percentage above 262,144, and those requests either fail, need truncation, or move to the Max route at more than double the input rate. Price the tail separately. If truncation is acceptable, say so explicitly in the product contract; if it is not, budget the higher route for the tail.
Cache reads only count when the prefix is stable
Max lists a $0.25 cache-read rate against a $2.00 prompt rate, an 87.5% reduction on cached tokens. The discount is real but narrow: it applies to the prefix the provider recognizes, and it requires that prefix to remain stable across requests. System prompts, tool schemas, and long instructions are good candidates. Per-user documents and conversation history are not, unless the provider's cache semantics explicitly cover them.
Instrument the cache-read field on responses before you forecast. If cached tokens are near zero after a week of production traffic, remove the assumption from the budget and revisit the prompt structure. A cache discount that never materializes is worse than no discount, because it hides the true unit cost behind a plausible-looking model.
Worked estimate: 100,000 requests
At 100,000 requests per month, 1,500 input tokens and 350 output tokens, the Max route bills 150M input tokens at $2.00 ($300.00) plus 35M output tokens at $6.00 ($210.00), for $510.00 per month and $5.10 per thousand requests. On Prime the identical workload is $600.00 plus $420.00, or $1,020.00. On Flash it is $22.50 plus $16.45, or $38.95 — a 26x spread inside one model family.
That spread is the reason a family-level 'Qwen pricing' answer is unusable for a budget. The routes do not deliver the same capability, latency, or context, so the comparison is only meaningful after a quality gate. Run a bounded evaluation set through the candidate routes, record pass rates and failure modes, and only then compare cost per accepted task rather than cost per request.
Cross-provider comparison needs matched units
APIDir's accepted observation for Qwen on Together AI is $2.50 input and $6.25 output, checked August 22, 2026, against OpenRouter's $2.00/$6.00. The gap is small enough that convenience, latency, rate limits, and data terms should decide, not a 20% headline difference. Two channels also do not prove infrastructure independence: if both resolve to the same upstream, a provider incident affects both.
When you compare, match the model version, the context ceiling, the cache terms, and the processing tier. A dated snapshot such as qwen3.8-max-0902 is comparable to another dated snapshot; a moving alias is not. Record the exact identifier in your configuration and in the invoice sample you reconcile against, or the comparison will drift as the catalog changes.
Budget controls that fit a Qwen workload
Cap output tokens per request, cap retrieval depth, and monitor the share of requests above the batch context ceiling. Those three controls address the meters that move a Qwen bill. Add a cache-hit dashboard only if the cache-read field is actually populated; otherwise the panel reports a number that never changes and gives false comfort.
Finally, recheck the route price before each planning cycle. Model catalogs change rates, context limits, and variants without renaming the family, and a budget built on a retrieved rate is only as good as its date. The table above carries the retrieval date for exactly this reason.
Price one Qwen route on your own traffic
Use the rates from the exact route row you plan to deploy. The estimate separates uncached input, cache-hit input, output, and batch share.
Planning result
Review the example inputs, replace each rate with the exact current value from the linked source, then calculate. APIDir does not send these values to a provider.
Questions developers ask
How much does Qwen3.8 Max cost per million tokens?
OpenRouter listed $2.00 per million input tokens and $6.00 per million output tokens on November 13, 2026, with a $0.25 cache-read rate. Treat that as a retrieved observation and recheck the route before you budget.
Is the Qwen batch route always cheaper?
No. It is cheaper per token but caps context at 262,144 tokens, so any request above that window must use a different route or be restructured.
What does the 1M context window mean for cost?
It is a ceiling, not a target. Long prompts multiply the input meter, and a single oversized request can dominate a month of short ones.
Should I compare Qwen on OpenRouter with Together AI?
Yes, but match the model version, context ceiling, cache terms, and processing tier first. The accepted observation gap on APIDir is $2.00/$6.00 versus $2.50/$6.25.
Sources and check dates
Every rate in this guide was retrieved from the linked page on Nov 13, 2026. Prices can change; recheck the source before purchasing.