Qwen3.8-Omni-Flash API Pricing
Qwen3.8-Omni-Flash is the Qwen3.8 model for listening and watching: it takes text, images, audio and video and answers in text, over a 1M-token context. Alibaba's International list prices it at $0.15/M input, $0.016/M on a cache hit and $0.47/M output, with one input rate for every modality; the Qwen3.5-Omni models charge audio input separately and several times higher. Audio counts 7 tokens per second, so an hour-long recording is 25,200 input tokens, well under a cent. It does not speak: for spoken replies Alibaba points to its realtime sibling or Qwen3.5-Omni-Plus.
Run the numbers.
Calculator pre-loaded with the International (Singapore) list rates. To size audio or video, convert it to tokens first: 7 tokens per second of audio, and for images one token per 32x32 pixels up to 1,280 tokens each. Thinking is on by default and its tokens bill as output.
Real-world presets.
One-hour call summary
Episode show notes
Voice-note ticket triage
Screenshot walkthrough
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.
| Model | Input /M | Output /M | Effective blended | Context | Best for |
|---|---|---|---|---|---|
| Qwen3.8-Omni-Flash Current | $0.15 cache $0.016 | $0.47 | $0.0745 agentic 92/8 | 1M | Audio and video understanding at a text-model rate |
| Qwen3.8-Flash | $0.15 | $0.47 | $0.176 +2.4× | 1M | The same rate for text and image work |
| Qwen3.8 Max | $2.00 | $6.00 | $2.32 +31× | 1M | Qwen's frontier tier for harder reasoning |
| MiMo-V2.6-Flash | $0.14 cache $0.0028 | $0.28 | $0.0477 −36% | 1M | Omni-modal input at a lower output rate |
| Gemini 3.8 Flash | $0.75 cache $0.075 | $3.75 | $0.481 +6.5× · promo | 1M | Multimodal input from a hyperscaler stack |
| GPT-6 Luna | $0.10 cache $0.01 | $0.50 | $0.0682 −8% | 1.05M | High-volume work on a closed frontier stack |
Frequently asked.
What Qwen3.8-Omni-Flash costs, how audio, images and video turn into tokens, and what it cannot do.
Q · 01 How much does Qwen3.8-Omni-Flash cost? +
Q · 02 How much does an hour of audio cost? +
$0.0038 of input at list. The answer is billed separately as output: a 3,000-token summary adds about $0.0014. Transcribing or summarising a recorded meeting therefore costs a fraction of a cent, and the text you ask for back usually costs more than the audio you send.Q · 03 How are images and video counted? +
$0.0002 by default. Setting vl_high_resolution_images raises the ceiling to 16,384 tokens. A video is billed as two streams: its sampled frames, counted like images, and its audio track at 7 tokens per second.Q · 04 Can it reply with speech? +
qwen3.8-omni-flash-realtime, released on September 21 with its own price card, and non-real-time audio output to qwen3.5-omni-plus. Use this model when the output you need is a summary, transcript, answer or structured data.Q · 05 Why use it instead of Qwen3.8-Flash? +
$0.016, which Qwen3.8-Flash does not, so on repeated long context it can come out cheaper per task.Q · 06 How does it compare with other omni-modal models? +
$0.14 input, $0.0028 cache hit and $0.28 output, so it is cheaper on every line and has open weights. Qwen3.8-Omni-Flash's case is Alibaba's documented audio handling: 113 languages and dialects, stereo and four-channel spatial audio, and a published per-second token rule you can budget against.Q · 07 What else is included? +
xhigh effort and can be turned down or off; thinking tokens bill at the output rate. Function calling and web search are supported. Implicit caching and Responses session caching are where the cache-hit rate applies. There is no batch discount and no free quota on this row.