DataLLM LabQwen3.8-Max: Alibaba Raised the Price 36% and the Bill Went Down
Qwen3.8-Max lists at $2.00 per million input tokens and $6.00 output. The model it replaces, Qwen3.7-Max, listed at $1.48 and $4.42 when we measured it on 2026-07-30. That is a 36% price increase on both sides, and by the logic of every pricing comparison table ever published, an upgrade that costs you more. We ran both generations on the same nine executed Python tasks. Both scored 9 out of 9. The measured cost went from $6.10 to $4.46 per 1,000 tasks — down 27%. The reason is not a discount. Qwen3.8-Max emitted 6,363 output tokens where 3.7-Max emitted 12,193. It got twice as concise, and conciseness outran the price rise.
Generational comparisons usually come from the vendor. This one comes from running both generations through the identical harness eight weeks apart, which is the only way to see the number that actually decides your invoice.
Two generations, same nine tasks
| Qwen3.7-Max | Qwen3.8-Max | Change | |
|---|---|---|---|
| List price in / out | $1.48 / $4.42 | $2.00 / $6.00 | +36% |
| Score | 9/9 | 9/9 | same |
| Measured cost / 1k tasks | $6.10 | $4.46 | −27% |
| Median latency | 25.8s | 16.8s | −35% |
| Output tokens | 12,193 | 6,363 | −48% |
| Reasoning tokens | 1,236 | 589 | −52% |
| Measured on | 2026-07-30 | 2026-08-20 | — |
Every column except price moved in your favour, and the price column is the only one most buyers ever see.
Why a 36% price rise made it cheaper
Your bill is price multiplied by volume. Providers publish the first factor and never the second, so a table of list prices can point the wrong way — as it does here.
Half the saving is reasoning tokens: 1,236 down to 589. Reasoning bills at the output rate, you cannot see it, and on these tasks the extra thinking bought nothing — both generations solved all nine. It is the same variable that dominates agent bills in our agent model comparison, and we saw it run the other way this week: GPT-5.1-Codex-Max has a lower list price than GPT-5.2-Codex and costs 3.09x more, for the same reason in reverse.
The rest of the 3.8 family
Qwen shipped a larger flagship alongside Max. It is not better on our tasks — it is shaped differently.
| Model | List price | Score | Measured cost / 1k | Latency | Reasoning tokens |
|---|---|---|---|---|---|
| Qwen3.8-Max | $2.00 / $6.00 | 9/9 | $4.46 | 16.8s | 589 |
| Qwen3.8-2.4T-A95B | $2.00 / $6.00 | 9/9 | $5.66 | 9.7s | 808 |
Identical list price, identical score. The 2.4T flagship is 27% more expensive and 42% faster. If latency is what you are buying, it is the one to pick; if the invoice is, it is not. Neither is a general upgrade on the other, which is a more useful thing to know than a leaderboard position.
The one we could not finish
We also ran Qwen3.8-27B. It solved the first eight tasks, then parse_csv_line failed at the API layer after retries — two empty responses with finish_reason=length, then a timeout. That is a run of eight scored tasks, not nine.
We have marked it excluded in our data rather than publishing 8/8 next to a column of 9/9 results, because the denominators are not the same and putting them side by side would imply a comparison that does not exist. This is the same call we made on an earlier Qwen model that hit the identical failure. If you want the reasoning behind treating harness failures as harness failures rather than model failures, the content filter post covers the time this distinction nearly cost a model five points.
Where this sits against the field
At $4.46 per 1,000 tasks, Qwen3.8-Max is not a budget option. Qwen3 Coder Next, from the same vendor, scored 9/9 at $0.10 in 7 seconds — roughly 45x cheaper on the same tasks, though with a 262,144-token context against Max's 1,000,000.
That gap is the real decision. If you need the million-token window or whatever Max does that nine Python functions cannot reveal, the price is the price. If you are picking a coding model on cost per finished task, the cheap end of Qwen's own lineup wins by a wide margin, and the cheap coding roundup and open-source guide cover the alternatives.
How these numbers were produced
Nine Python tasks, each a function signature plus a spec and no example tests. Generated code executes against hidden asserts in an isolated python3 -I subprocess with a 12-second timeout. Temperature 0, max_tokens 4000, one scored attempt per task. Cost is derived from measured token counts at the list price on the date shown, not a billing statement — and since the two generations were priced three weeks apart, price movement is exactly why both dates are on the table above. Runs go through OpenRouter. Full method on the methodology page.
What we did not measure
- The million-token context. Our prompts are short. Nothing here tests long-context behaviour, which is a large part of what Max is sold on.
- Anything but Python. The suite is Python only, and nine self-contained functions at that.
- Repeat runs. One scored attempt per task per model. Single-run figures, not averages, and we do not average across runs.
- Qwen3.8-27B completely — eight of nine tasks, see above.
- The open-weight 3.8 variants. The Uncensored, Abliterated-SFT and 2B-Distill releases are separate models and are not covered by anything here.
FAQ
Is Qwen3.8-Max cheaper than Qwen3.7-Max? Per token, no — it lists 36% higher. Per finished task, yes: $4.46 against $6.10 on our nine tasks, a 27% reduction.
How much did Qwen3.8-Max improve? Same score, 35% lower latency, 48% fewer output tokens and 52% fewer reasoning tokens than Qwen3.7-Max on identical prompts.
Qwen3.8-Max or Qwen3.8-2.4T-A95B? Same list price and same 9/9. The 2.4T is 42% faster and 27% more expensive per task. Pick on which of those you care about.
What happened to Qwen3.8-27B? One task failed at the API layer after retries, so it completed eight of nine. We excluded it rather than report a score with a different denominator.
Is Qwen3.8-Max good value for coding? Not against its own cheap sibling. Qwen3 Coder Next scored the same 9/9 at $0.10 per 1,000 tasks.