QWEN 3.8 FLAGSHIPOUT OF PREVIEW1M CONTEXTTEXT + VISION + VIDEOTHINKING SAME PRICE
Qwen3.8 Max API Pricing
Qwen3.8 Max is Alibaba's newest Qwen Max flagship, and as of August 2026 it finally has a price. The International / Singapore row lists $2/M input and $6/M output for 0-1M tokens, published directly in US dollars. That undercuts Qwen3.7 Max, the flagship it replaces, on both sides of the meter — the newer model is the cheaper one. Thinking and non-thinking modes bill at the same rate.
Input - per 1M tokens
$2.00/M
Base token price standard
Output - per 1M tokens
$6.00/M
Output tokens standard
Cached input - not itemized
$2.00/M
Cache discount noted, price not itemized not itemized
Effective - agentic blend
$2.32/M
92/8 split - no cache discount
§ 01 / TERMINAL
Qwen3.8 Max API cost calculator.
Calculator pre-loaded with the International Qwen3.8 Max rates. Alibaba publishes no per-model cache rate, so the cache slider stays off — what you see is the undiscounted cost.
$ /mo
Workload split
Tokens you can process
—
Words equivalent (English)
—
Effective rate
—
§ 02 / SCENARIOS
Qwen3.8 Max cost on real workloads.
CODING
Repo-wide feature build
$0.300/task
LONG DOC
Reading 150-page contracts
$0.408/doc
RAG
Support agent ticket triage
$0.022/ticket
ASSISTANT
Research planning turn
$0.048/turn
Calibrated · measured on the vendor's tokenizer · 2026-06-10 Auto-counts as you type
Counts use a chars-per-token calibration measured on the vendor's own published tokenizer (Qwen/Qwen3.5-397B-A17B, 2026-06-10). English prose is typically within a few percent; code and non-Latin scripts tokenize heavier. For billing-exact counts use the vendor's count-tokens API.
Characters 407
Words 64
Tokens (estimated) 78 tokens
Cost as input · uncached $0.00016 USD
Cost as output · uncached $0.00047 USD
| Model | Input /M | Output /M | Effective blended | Context | Best for |
|---|---|---|---|---|---|
| Qwen3.8 Max Current | $2.00 | $6.00 | $2.32 agentic 92/8 | 1M | Newest Qwen flagship, now multimodal |
| Qwen3.7 Max | $2.50 | $7.50 | $2.90 +25% | 1M | Previous Max flagship, text only |
| Qwen3.7 Plus | $0.40 | $1.60 | $0.496 −4.7× · promo | 1M | Cheaper Qwen tier for bulk work |
| GPT-5.6 Terra | $2.00 cache $0.20 | $12.00 | $1.52 −34% | 1.05M | Same input rate, real cache discount |
| Claude Sonnet 5.5 | $2.00 cache $0.10 | $10.00 | $1.29 −44% | 1M | Anthropic's current Sonnet |
| DeepSeek V4 Pro | $1.32 cache $0.044 | $3.96 | $0.569 −4.1× | 1M | Cheapest 1M reasoning tier |
Qwen3.8 Max has published scores on 12 benchmarks we track, and leads SWE-Bench Pro. 1 of these scores comes from an independent leaderboard; the rest are the lab's own reports.
| Benchmark | Score | Rank | Leader | Source |
|---|---|---|---|---|
| Agents' Last Exam Agents and computer use · Strict pass rate | 27% pass rate | #2 of 5 | Gemini 4 Argon · 39.5% | Vendor-reported Alibaba (Qwen) |
| Terminal-Bench 2 Agents and computer use · Terminal-Bench 2.1 | 86.6% as reported | #8 of 17 | DeepSeek V4.1 Flash · 90.6% | Vendor-reported Alibaba (Qwen) |
| SWE-Bench Pro Coding · Set not stated | 67.7% set not stated | #1 of 2 | Qwen3.8 Max leads | Vendor-reported Alibaba (Qwen) |
| GPQA Diamond Expert knowledge | 92.6% as reported | #6 of 11 | GPT-6 Astra · 96% | Vendor-reported Alibaba (Qwen) |
| FrontierSWE Coding · Version not stated | 73.5% as reported | #3 of 5 | Kimi K3 · 81.2% | Vendor-reported Alibaba (Qwen) |
| Humanity's Last Exam Expert knowledge · With tools | 56.2% with tools | #12 of 20 | Claude Fable 5.2 · 74.2% | Vendor-reported Alibaba (Qwen) |
| Terminal-Bench 4.0 Agents and computer use | 27% max effort, Claude Code harness, 330 trials, ±3.8 pts (95% CI) | #21 of 34 | Kimi K3.1 Pro · 65.7% | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| NL2Repo-Bench Coding | 55.9% as reported | #5 of 8 | DeepSeek V4.1 Flash · 68.2% | Vendor-reported Alibaba (Qwen) |
| Toolathlon Agents and computer use | 72.5% pass@1 | #6 of 9 | MiMo-V2.6-Pro · 76.9% | Vendor-reported Alibaba (Qwen) |
| DeepSWE Coding | 56.6% as reported | #20 of 27 | Gemini 4 Argon · 77.9% | Vendor-reported Alibaba (Qwen) |
| AutomationBench Agents and computer use · Main set (as reported) | 27.3% pass@1 | #17 of 22 | MiniMax M3.1 Flash · 62.8% | Vendor-reported Alibaba (Qwen) |
| JobBench Professional work | 53.4% as reported | #4 of 5 | MiMo-V2.6-Pro · 62% | Vendor-reported Alibaba (Qwen) |
Qwen3.8 Max API pricing FAQ.
Short answers for teams checking Qwen3.8 Max pricing, regional rate cards, and what changed when it left preview.
Q · 01 What is Qwen3.8 Max priced at? +
Alibaba's International (Singapore) rate card lists $2/M input and $6/M output, published directly in US dollars, in a single billing band covering 0-1M input tokens per request. Under AI//COST's 92/8 agentic blend the effective planning figure is $2.32/M — there is no cache discount to apply, so that is simply the undiscounted mix.
Q · 02 Why do I see $1.65 and $4.951 elsewhere? +
Because Alibaba publishes several rate cards on one page and they are different products.
$1.65/$4.951 is the Global / China deployment card; $2/$6 is the International (Singapore) card. AI//COST quotes the International list for every Alibaba model, so the numbers stay comparable across our catalogue. Mixing the two is exactly what produced the pricing error we corrected on July 27, 2026.Q · 03 Did the price change when it left preview? +
There was no price to change. From its WAIC announcement on July 19, 2026 the model was
qwen3.8-max-preview, reachable only through a Token Plan subscription in the Beijing region, with no per-token rate published anywhere. We ran the page with a blank price board rather than estimate one. The formal qwen3.8-max ID has now replaced it and carries the first real rate this model has ever had.Q · 04 Is it cheaper than Qwen3.7 Max? +
Yes, on both sides: $2 against $2.50 input, $6 against $7.50 output — about 25% cheaper on the blended agentic figure, and it adds image and video input that Qwen3.7 Max does not have. One caveat worth checking before you migrate: Qwen3.7 Max currently carries a limited-time 50% discount on its list price, so what you are billed today may be lower than its list suggests.
Q · 05 Is there a Batch discount? +
Not on the International card. Alibaba states that batch calls bill at 50% of the real-time price for models marked as supporting them, and the International row for Qwen3.8 Max carries no batch marker — only a context-caching one. The Global/China rows for this model do show the batch marker. We record no batch rate rather than assume the discount carries across regions.
Q · 06 Is context caching priced? +
It is offered but not itemized. The rate card marks Qwen3.8 Max as having a context-caching discount without publishing a per-model cache-hit rate, so we leave the cached row at the standard input price and keep the calculator's cache slider off. Every figure on this page is therefore the undiscounted cost — your real bill with heavy caching will be lower by an amount Alibaba does not disclose.
Q · 07 Which API model ID should I use? +
qwen3.8-max. The preview ID qwen3.8-max-preview has been withdrawn from Alibaba's rate cards. The model is served from Beijing, Singapore, Tokyo, Frankfurt and Virginia endpoints, and Alibaba exposes OpenAI-compatible, Anthropic-compatible and DashScope interfaces for it.Q · 08 Does it accept images and video? +
Yes. Alibaba lists Qwen3.8 Max under both text generation and image/video understanding in its model catalogue, which is a change from Qwen3.7 Max — that tier is text-only. The rate card publishes a single token price covering the model rather than separate per-modality rates.