Qwen 3.8 Pricing Explained: $2 Input / $6 Output Across a 1M Context

Alibaba released Qwen 3.8-Max in early August 2026, and its general-availability pricing is now clear. The official Model Studio pricing page lists qwen3.8-max at $2 per 1M input tokens and $6 per 1M output tokens, with one flat tier covering the full 1M-token context window.

Try Apidog today

The flat tier is the key detail: whether you send 5,000 tokens or 900,000 tokens, the input rate remains $2 per 1M tokens.

This guide covers the rate card, caching, free-quota limits, comparisons with Qwen 3.7-Max, Kimi K3, Claude Opus 5, and GPT-5.6 Terra, plus worked cost calculations. For the broader model overview, start with what Qwen 3.8 is and why it matters.

Sticker prices are only an estimate. Qwen 3.8-Max defaults to xhigh reasoning effort, and thinking tokens are billed as output. Measure real requests before forecasting spend. Apidog exposes token usage for each response, making it easier to calculate actual per-task cost.

The GA rate: $2 input, $6 output, one tier

As of August 3, 2026, the official rate card for qwen3.8-max is:

Item	Price per 1M tokens
Input	$2.00
Output, including thinking tokens	$6.00
Cached input, cache hit	10% of the input rate
Explicit cache creation	125% of the input rate
Context tier	One tier: 0 to 1M tokens
Thinking vs. non-thinking	Same rates

Implementation implications:

• No long-context surcharge: The 1M-token context window uses one price tier.
• Thinking tokens are output: Reasoning increases output token usage and therefore cost.
• Maximum output is 65,536 tokens: At $6 per 1M output tokens, a maximum-length response costs about $0.39 in output tokens.

The official Qwen 3.8 announcement describes the model as Alibaba’s most capable model yet, with 2.4T total parameters, 95B active parameters, a 1M-token context window, and multimodal input.

Why flat pricing across 1M tokens matters

Tiered context pricing is common: a vendor charges one rate below a context threshold and a higher rate above it.

Alibaba uses tiered pricing for other models. On the Model Studio pricing page, models such as qwen3-max and qwen3-coder-plus increase per-token pricing as context length grows.

Qwen 3.8-Max instead uses a single 0<Token≤1M tier.

For implementation planning, that makes the cost formula linear:

request_cost =
  (input_tokens / 1_000_000 × $2)
  + (output_tokens / 1_000_000 × $6)

This is especially useful for:

• Long-document analysis
• Large-codebase agents
• Retrieval-heavy RAG pipelines
• Workflows that sometimes send hundreds of thousands of tokens

With tiered pricing, an occasional 300K-token retrieval request can enter an expensive bracket. With Qwen 3.8-Max, each input token uses the same rate.

Qwen 3.8-Max vs. Qwen 3.7-Max

Qwen 3.8-Max launched below its predecessor’s list price:

• Qwen 3.7-Max: $2.50 input / $7.50 output per 1M tokens
• Qwen 3.8-Max: $2.00 input / $6.00 output per 1M tokens

That is a 20% reduction for both input and output.

However, Qwen 3.7-Max is running a 50% promotion:

• Qwen 3.7-Max promotional rate: $1.25 input / $3.75 output per 1M tokens

If you do not need Qwen 3.8-Max’s 1M context, multimodal input, or agentic improvements, discounted Qwen 3.7-Max can be a lower-cost choice while the promotion lasts.

For lower-cost workloads, Qwen 3.7-Plus remains a budget option at $0.40 input / $1.60 output, with its own 20% promotion.

Thinking tokens are billed as output

Qwen 3.8-Max supports three reasoning_effort levels:

xhigh
medium
low

The default is xhigh.

In thinking mode, the model generates internal reasoning tokens before returning its final answer. These tokens are billed as output tokens at $6 per 1M tokens.

That means this estimate can be misleading:

cost ≈ prompt tokens + visible response tokens

A visible 800-token answer may include several thousand additional thinking tokens on a difficult task.

Control reasoning cost by task type

Use lower reasoning effort for tasks such as:

• Classification
• Extraction
• Formatting
• Simple transformations
• Structured-output generation

For example, test multiple effort levels against representative requests:

{
  "model": "qwen3.8-max",
  "messages": [
    {
      "role": "user",
      "content": "Extract the invoice number, total, and due date as JSON."
    }
  ],
  "reasoning_effort": "low"
}

For harder planning, coding, or multi-step analysis tasks, benchmark medium and xhigh rather than assuming that the highest setting is always necessary.

Always inspect the response usage object. Output costs 3× more than input, and reasoning primarily increases output usage.

Context caching: 10% cache hits, 125% creation

Context caching is useful when requests repeatedly include the same prefix, such as:

• Long system prompts
• Product documentation
• Tool definitions
• Stable policy instructions
• Large reference documents

The pricing is:

• Cache hit: 10% of the normal input rate
  • $0.20 per 1M cached input tokens instead of $2.00
• Explicit cache creation: 125% of the normal input rate
  • $2.50 per 1M tokens

The break-even logic is straightforward:

• Creating a cache costs an extra 25% once.
• Each cache hit saves 90% relative to normal input pricing.
• Reusing a prefix twice is already enough for caching to pay off.

Free quota: 1M tokens, Singapore only, 90 days

Model Studio provides a free quota for qwen3.8-max, with important constraints:

• Quota: 1M tokens
• Region: Singapore only
• Expiry: 90 days from activation

Use the Singapore endpoint:

https://dashscope-intl.aliyuncs.com/compatible-mode/v1

The quota does not apply to Beijing or US-Virginia endpoints.

Thinking tokens consume this quota as output tokens, so a small number of difficult xhigh reasoning tasks can use the quota faster than expected.

For all no-cost access options, including Qwen Chat, see how to use Qwen 3.8 for free.

Pricing comparison

List prices per 1M tokens:

Model	Input	Output	Notes
Qwen 3.8-Max	$2.00	$6.00	Flat across 1M context; cache hits at 10%
Qwen 3.7-Max	$2.50	$7.50	Currently 50% off: $1.25/$3.75
Qwen 3.7-Plus	$0.40	$1.60	Currently 20% off
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	—

Key comparisons:

• Against Kimi K3: Qwen 3.8-Max costs one-third less on input and 60% less on output. Its cache-hit rate is also lower: $0.20 vs. $0.30. See the Kimi K3 API guide.
• Against Claude Opus 5: Qwen is 60% cheaper on input and 76% cheaper on output.
• Against GPT-5.6 Terra: Input pricing is the same at $2 per 1M tokens, but Qwen output is half the price.

Price is not a quality benchmark. Test models on your own workload before choosing one based on rate cards alone.

Worked cost examples

Use this formula:

input_cost = input_tokens × 2 / 1_000_000
output_cost = output_tokens × 6 / 1_000_000
total_cost = input_cost + output_cost

Example 1: 50K-token agent session

An agent session uses:

• 38,000 input tokens for instructions, tool schemas, and tool results
• 12,000 output tokens

Calculation:

Input: 38,000 × $2 / 1,000,000 = $0.076
Output: 12,000 × $6 / 1,000,000 = $0.072
Total: $0.148

Estimated cost: about $0.15 per session

Now include 8,000 thinking tokens from default xhigh reasoning:

Total output tokens: 12,000 + 8,000 = 20,000
Output: 20,000 × $6 / 1,000,000 = $0.12
Input: $0.076
Total: $0.196

Estimated cost with reasoning: about $0.20 per session

Reasoning increases this example by about 33%.

Example 2: 800K-token document analysis

You provide 800,000 input tokens and request 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: $1.63

Estimated cost: about $1.63 per run

The important part is that all 800K input tokens use the same $2-per-1M rate.

Example 3: Cached 100K-token prefix, 50 calls per day

Your application sends 100,000 tokens of system prompt, documentation, and tool definitions on every request.

Without caching:

100,000 × $2 / 1,000,000 = $0.20 per call
$0.20 × 50 calls = $10.00 per day

With explicit caching:

Cache creation:
100,000 × $2.50 / 1,000,000 = $0.25

49 cache hits:
100,000 × $0.20 / 1,000,000 = $0.02 per hit
$0.02 × 49 = $0.98

Daily total:
$0.25 + $0.98 = $1.23

Estimated daily cost: about $1.23 instead of $10.00

That is about an 88% reduction for the repeated prefix.

Build your cost model from measured usage

The calculations above use assumed token counts. Your actual counts will vary, especially when reasoning tokens are enabled.

Use this workflow:

1. Get an API key.
2. Configure the appropriate regional endpoint.
3. Send representative requests for every task type.
4. Record the response usage object.
5. Separate input, output, and reasoning-related output usage.
6. Run multiple tests per task type.
7. Average usage and apply the $2/$6 rates.

For endpoint setup and protocol options, see the Qwen 3.8 API guide, which covers regional base URLs plus OpenAI-compatible and Anthropic-compatible APIs.

In Apidog, save the regional endpoints as environments and send the same request with each reasoning_effort level. Inspect token usage directly in each response.

A practical test matrix might look like this:

Task type	Tested effort levels	What to measure
Classification	low, medium	Accuracy, output tokens
Extraction	low, medium	Schema compliance, output tokens
Coding	medium, xhigh	Test pass rate, total output
Agent workflows	medium, xhigh	Tool success rate, total cost
Long-context analysis	medium, xhigh	Quality, input and output usage

Run at least 10 representative requests for each task type. Then calculate:

monthly_cost =
  average_request_cost
  × requests_per_month

You can also A/B the same prompts against qwen3.7-max or Kimi K3 in one workspace. Download Apidog to capture real per-task token usage while testing.

Bottom line

Qwen 3.8-Max is priced at $2 per 1M input tokens and $6 per 1M output tokens, below Qwen 3.7-Max’s list price and with a flat rate across the full 1M-token context window.

For long-context or high-repetition workloads, the combination of flat input pricing and 10% cache-hit pricing can be compelling.

Before committing to a production cost estimate:

• Measure real token usage.
• Treat thinking tokens as output costs.
• Set reasoning_effort deliberately.
• Use caching for repeated prompt prefixes.
• Do not build long-term forecasts around time-limited promotions or free quotas.

Frequently asked questions

Does Qwen 3.8 cost more for long prompts?

No. The official pricing shows one tier covering 0 to 1M tokens. A 900K-token prompt uses the same $2-per-1M input rate as a short prompt.

This differs from tiered models, including models listed in Alibaba’s Model Studio model list, where rates can increase with context length.

Do thinking tokens cost extra on Qwen 3.8?

There is no separate thinking-token rate. Thinking and non-thinking requests use the same $2 input / $6 output rates.

However, thinking tokens count as output tokens, and reasoning_effort defaults to xhigh. Reasoning-heavy requests can therefore cost more than the visible answer suggests.

Use lower reasoning effort for simple tasks and inspect the response usage object for actual token counts.

Is there a free way to try Qwen 3.8?

Yes. Model Studio includes a 1M-token free quota valid for 90 days, but only on the Singapore endpoint. Qwen Chat also provides no-cost browser access.

See how to use Qwen 3.8 for free for the available options and limitations.

Is the old “10% preview pricing” still available?

No. The 10% pricing was a July 2026 preview-era promotion. General availability uses the standard $2/$6 rate card. Use the official Model Studio pricing page as the source of truth.