Alibaba released Qwen 3.8-Max in early August 2026, and the pricing question finally has a real answer. During the July preview window, the story circulating was a “10% preview pricing” promotion. That story is over. At general availability, the official Model Studio pricing page lists qwen3.8-max at $2 per 1M input tokens and $6 per 1M output tokens, one flat tier covering the entire 1M-token context window.
That flat tier is the interesting part. Most frontier models charge more per token once your prompt crosses a context threshold. Qwen 3.8-Max doesn’t: send 5,000 tokens or 900,000, the rate stays $2/$6.
This article covers the full picture: the flat-tier design, comparisons against Qwen 3.7-Max, Kimi K3, Claude Opus 5, and GPT-5.6 Terra, the caching discounts, the free quota fine print, and worked cost examples with real math. For the broader model overview, start with what Qwen 3.8 is and why it matters, then come back for the numbers.
One note first: sticker prices only get you an estimate. Qwen 3.8-Max ships with reasoning at xhigh effort by default, and thinking tokens bill as output. The reliable way to project costs is to send real requests and measure token usage. Apidog shows the full token breakdown of every response while you test, which turns cost estimation into arithmetic. More on that below.
The GA number: $2 input, $6 output, one tier
Here’s the official rate card for qwen3.8-max, verified against the Model Studio pricing page as of August 3, 2026:
| Item | Price (per 1M tokens) |
|---|---|
| Input | $2.00 |
| Output (including thinking tokens) | $6.00 |
| Cached input (cache hit) | 10% of input rate |
| Explicit cache creation | 125% of input rate |
| Context tier | Single tier, 0 to 1M tokens |
| Thinking vs non-thinking | Billed the same |
Three details worth calling out: the 1M context window is fully covered by one price, with no long-prompt surcharge. Thinking and non-thinking modes cost the same rates, though thinking tokens count as output (see the honesty section below). And max output is 65,536 tokens per request, so a single response bills at most about $0.39 in output.
The official Qwen 3.8 announcement positions the model as Alibaba’s most capable yet: 2.4T total parameters with 95B active, a 1M-token context window, multimodal input. Getting that at $2/$6 undercuts most of the frontier tier, including Alibaba’s own previous flagship.
Why flat pricing across 1M tokens is unusual
Tiered context pricing is the industry norm: one rate below a context threshold, a higher rate above it, because long contexts cost more to serve.
Alibaba does this itself. On the same Model Studio pricing page, models like qwen3-max and qwen3-coder-plus are priced in context-length tiers, with the per-token rate stepping up as input grows. Qwen 3.8-Max breaks from that pattern within Alibaba’s own catalog: its pricing table shows one row, “0<Token≤1M”, and that’s it.
Why this matters: with tiered pricing, long-context workloads carry a hidden multiplier. A RAG pipeline that occasionally stuffs 300K tokens of documents into a prompt can cost several times its typical rate on those requests. With Qwen 3.8-Max, the marginal token always costs the same, so budgeting is linear: tokens times rate, done. For long-document analysis, large-codebase agents, or heavy retrieval pipelines, that predictability is worth as much as the low rate itself.
Cheaper than Qwen 3.7-Max at list price
Qwen 3.8-Max launched below its predecessor’s list price, which is rare for a flagship generation jump:
- Qwen 3.7-Max list price: $2.5 input / $7.5 output per 1M tokens
- Qwen 3.8-Max: $2 input / $6 output per 1M tokens
That’s a 20% cut on both sides while the model gained a bigger context window, multimodal input, and better benchmark scores.
One complication: Qwen 3.7-Max is currently running a 50% promotion, which brings it down to $1.25/$3.75. At promo rates, the older flagship is cheaper than the new one. If your workload doesn’t need Qwen 3.8’s 1M context or its multimodal and agentic gains, discounted 3.7-Max is a legitimate value play while the promotion lasts; just don’t architect your cost model around a promo. Further down the ladder, Qwen 3.7-Plus remains the budget workhorse at $0.4/$1.6, with a 20% promotion of its own.
The honesty section: thinking tokens bill as output
This is the part most pricing writeups skip, and the part most likely to surprise you on an invoice.
Qwen 3.8-Max supports a reasoning_effort parameter with three levels: xhigh, medium, and low. The default is xhigh, the most aggressive setting. In thinking mode, the model generates internal reasoning tokens before its final answer, and those reasoning tokens bill at the $6 output rate, same as visible text.
The consequence: at default settings, reasoning-heavy requests cost more than a naive “prompt tokens plus answer tokens” estimate predicts. A response showing 800 tokens of visible answer may have consumed several thousand thinking tokens on a hard problem.
Three mitigations: lower reasoning_effort to medium or low for tasks that don’t need deep reasoning (classification, extraction, formatting); measure actual usage per task type before projecting monthly spend, since the usage object in every response reports real counts including reasoning; and watch output tokens specifically, because output costs 3x input here and thinking inflates output.
Context caching: 10% hits, 125% creation
If your application resends the same prompt prefix repeatedly (a long system prompt, a stable document, tool definitions), context caching changes the economics significantly:
- Cache hits bill at 10% of the input rate. Cached input tokens cost $0.20 per 1M instead of $2.
- Explicit cache creation bills at 125% of the input rate. Writing tokens into the cache costs $2.50 per 1M, a one-time 25% premium over normal input.
The break-even math is simple. Creation costs an extra 25% once; every subsequent hit saves 90%. If a prefix gets reused even twice, caching already pays for itself. For high-frequency agents or chat applications with a heavy system prompt, the savings compound fast (worked example below).
Free quota fine print: 1M tokens, Singapore, 90 days
Model Studio offers a free quota for qwen3.8-max, and the fine print matters:
- Size: 1M tokens. Enough for real evaluation, not for production.
- Region: Singapore only. Point your client at
https://dashscope-intl.aliyuncs.com/compatible-mode/v1; the quota doesn’t apply to Beijing or US-Virginia. - Duration: 90 days from activation. Unused quota expires.
Thinking tokens draw from this quota like any other output, so a few dozen hard reasoning tasks at xhigh can drain 1M tokens faster than you’d expect. If free access is your main goal, we’ve collected every no-cost route, including Qwen Chat, in how to use Qwen 3.8 for free.
How Qwen 3.8 pricing compares
Here’s the current landscape, using list prices per 1M tokens. Promotional rates are flagged, because promos expire.
| Model | Input | Output | Notes |
|---|---|---|---|
| Qwen 3.8-Max | $2.00 | $6.00 | Flat across 1M context; cache hits 10% |
| Qwen 3.7-Max | $2.50 | $7.50 | Currently 50% off: $1.25/$3.75 (promo) |
| Qwen 3.7-Plus | $0.40 | $1.60 | Currently 20% off (promo) |
| Kimi K3 | $3.00 | $15.00 | Cache hits $0.30 |
| Claude Opus 5 | $5.00 | $25.00 | |
| GPT-5.6 Terra | $2.00 | $12.00 |
Reading the table:
- Against Kimi K3, its closest open-weight rival, Qwen 3.8-Max costs a third less on input and 60% less on output, and its cache-hit rate ($0.20) undercuts K3’s ($0.30). For output-heavy agentic work, that output gap dominates. See the Kimi K3 API guide for K3’s side.
- Against Claude Opus 5, Qwen is 60% cheaper on input and 76% cheaper on output.
- Against GPT-5.6 Terra, input is identical at $2, but Qwen’s output is half the price. For long-form generation and reasoning-heavy work, that’s the number that moves your bill.
Price isn’t quality, and vendor benchmark tables are vendor benchmark tables. But at pure rate-card level, Qwen 3.8-Max is the cheapest frontier-class flagship in this lineup.
Worked examples: what real tasks cost
Sticker rates mean little without task-shaped math. Three scenarios, computed at $2/$6:
Example 1: a 50K-token agent session
An agent run consumes 38,000 input tokens (instructions, tool schemas, tool results) and produces 12,000 output tokens:
- Input: 38,000 × $2 / 1,000,000 = $0.076
- Output: 12,000 × $6 / 1,000,000 = $0.072
- Total: about $0.15 per session
Now add default xhigh reasoning. Suppose the model spends 8,000 thinking tokens across the run. Output becomes 20,000 tokens: 20,000 × $6 / 1,000,000 = $0.12, and the session totals about $0.20. That’s a 33% increase from reasoning alone, which is exactly why you measure before you budget.
Example 2: long-document analysis at 800K tokens
You load 800,000 tokens of documents into the context and ask for a 5,000-token structured summary:
- Input: 800,000 × $2 / 1,000,000 = $1.60
- Output: 5,000 × $6 / 1,000,000 = $0.03
- Total: about $1.63 per run
The flat tier is doing the work here. Every one of those 800K tokens costs the same $2/1M rate as the first thousand. On a tiered model, this is precisely the request shape that triggers the expensive bracket.
Example 3: a cached 100K-token system prompt, 50 calls per day
Your application front-loads 100,000 tokens of system prompt, product docs, and tool definitions on every call, 50 times a day.
Without caching: 100,000 × $2 / 1,000,000 = $0.20 per call, so $10.00 per day on the prefix alone.
With explicit caching: creation costs 100,000 × $2.50 / 1,000,000 = $0.25 once, then 49 cache hits at 100,000 × $0.20 / 1,000,000 = $0.02 each, or $0.98. Daily total: about $1.23, an 88% saving. Over a month, that’s roughly $300 versus under $40.
Estimate costs from measured usage, not guesses
All three examples rest on assumed token counts. Your real numbers will differ, and with reasoning tokens in play, they’ll differ in ways you can’t predict from prompt length alone. The fix is to measure.
A practical workflow: grab your API key, set up the endpoint (the Qwen 3.8 API guide covers all three regional base URLs and both the OpenAI-compatible and Anthropic-compatible protocols), and send representative requests for each task type in your application. Every response returns a usage object with exact input, output, and reasoning token counts.
In Apidog, save the three regional base URLs as environments, fire the same request at each reasoning_effort level, and read the token usage straight from the response inspector. Run ten representative requests per task type, average the counts, multiply by $2/$6, and you have a cost model built on measured data. You can also A/B the same prompt against qwen3.7-max or Kimi K3 in the same workspace to see whether the cheaper model holds up on your workload. Download Apidog for free and you can have real per-task cost numbers within an hour.
The bottom line
Qwen 3.8-Max at $2/$6 is aggressive pricing for a frontier-class flagship: below its predecessor’s list price, half of GPT-5.6 Terra’s output rate, a fraction of Opus 5, and flat across a full 1M-token context where much of the market tiers up. Add 10% cache hits and the economics for long-context, high-repetition workloads look genuinely strong.
Two caveats: default xhigh reasoning inflates output billing above naive estimates, and the anchors around it (3.7-Max at 50% off, the Singapore free quota) are time-limited. Measure your real token usage, set reasoning_effort deliberately, and the rate card holds few surprises.
Frequently asked questions
Does Qwen 3.8 cost more for long prompts?
No. The official pricing lists a single tier covering 0 to 1M tokens, so a 900K-token prompt bills at the same $2 per 1M input rate as a short one. This differs from tiered models, including some in Alibaba’s own catalog on the Model Studio model list, where rates step up with context length.
Do thinking tokens cost extra on Qwen 3.8?
There’s no separate thinking rate: thinking and non-thinking modes bill at the same $2/$6. But thinking tokens count as output tokens at $6 per 1M, and reasoning_effort defaults to xhigh. Reasoning-heavy requests therefore cost more than the visible answer suggests. Lower the effort level for simple tasks, and check the usage object in each response to see what you’re actually paying for.
Is there a free way to try Qwen 3.8?
Yes. Model Studio includes a free quota of 1M tokens, valid for 90 days, on the Singapore region endpoint only. Qwen Chat also offers no-cost access in the browser. Both routes, plus the fine print, are covered in how to use Qwen 3.8 for free.
Is the old “10% preview pricing” still available?
No. That was a preview-era promotion reported in July 2026 press coverage; general availability replaced it with the standard $2/$6 rate card. Treat the Model Studio pricing page as the source of truth.
