Alibaba Launches Qwen3.8-Omni-Flash Audio-Video Agent API at $0.15 Per Million Tokens

Alibaba's Qwen team launched Qwen3.8-Omni-Flash on September 18, 2026 — a multimodal model that takes text, images, audio, and video in a single API call, plans tasks across them, executes tools independently, and returns text — at $0.15 per million input tokens. That price sits at one-fifth of the current introductory rate for Gemini 3.8 Flash, and less than one-tenth of what Google's Flash pricing becomes on January 1, 2027, when introductory rates on the Gemini 3.x Flash line double. The model is live through QwenCloud, Alibaba Cloud Model Studio, and Qwen Studio, with no open weights at launch.

Read more: DeepSeek V4 Flash Vision Exp opens multimodal agent workflows at text-model prices

Qwen3.8-Omni-Flash Is Not a Perception Model — It Is an Execution Model

The design distinction Alibaba emphasizes is worth examining. Most multimodal models, when given a two-hour video and a question, process every frame the same way — scanning linearly and hoping the relevant content falls within the effective attention range. Qwen3.8-Omni-Flash takes a different approach Alibaba calls agentic perception: it starts from the user's question, determines which segments of the video and audio stream are worth examining, and gathers evidence in iterative coarse-to-fine rounds rather than processing the entire input at equal resolution.

On OmniVideoBench — a benchmark testing models against long video — the company reports that its agentic approach scores 67.8 against the 63.4 its static-processing mode achieves, while consuming 45.7% fewer tokens (79,117 versus 145,736). Those numbers are vendor-reported and have not been independently replicated at publication, but the underlying logic is sound: targeted, question-driven video reading scales better to very long content than uniform dense processing.

The practical payoff for developers is that the same 1 million-token context window that seems expensive when used to process footage linearly becomes tractable when the model learns to allocate tokens only to the portions of footage that answer the question.

Beyond understanding, the model can trigger tools on what it finds. A single API call can watch a meeting recording, identify the three decisions made, extract action items attributed to named speakers, and populate a task management system — without requiring a separate orchestration layer to chain those steps. Thinking mode is on by default, set to the highest reasoning effort (xhigh), and can be disabled for latency-sensitive deployments.

The Architecture Underneath: A Qwen4 Preview With 6 Billion Active Parameters

Qwen3.8-Omni-Flash is built on Qwen3.8-Flash-Next, the experimental open-weight model Alibaba released in late August 2026 as an early preview of the architecture it intends to use for Qwen4.

That base is a mixture-of-experts model with approximately 125 billion total parameters but only 6 billion active per token. The sparse activation is deliberate: most of a model's stored knowledge sits in expert layers that are never invoked for any given token, which means the actual compute per token is equivalent to running a 6-billion-parameter dense model while still drawing on the pattern knowledge encoded across all 125 billion. This is how inference at a $0.15-per-million-token price becomes viable for a model with genuine frontier-adjacent capabilities.

The key architectural change versus prior Qwen generations is Qwen Sparse Attention (QSA), which operates at the micro-block level rather than per-token. This reduces latency on long contexts significantly compared to conventional sliding-window attention — a meaningful gain when the model is processing 45-minute audio files or two-hour video in a single context. The Gated Residual mechanism and n-gram embedding vocabulary round out the architectural changes the team frames as a "systematic upgrade" on the path to Qwen4.

On top of the Flash-Next text-and-image base, the omni-modal variant adds two components: a vision encoder and a dedicated audio-text module. Both feed into the same central agent that plans actions and calls tools. Critically, this is an end-to-end native architecture rather than a pipeline — the model's attention layers see all four modalities simultaneously rather than receiving pre-converted representations. The distinction matters for tasks where meaning is distributed across channels simultaneously, such as a speaker's tone of voice contradicting the literal content of what they say.

One limitation is conspicuous: Qwen3.8-Omni-Flash accepts audio and video as input but produces text only as output. Teams that need generated speech must chain the model with a separate text-to-speech step; Alibaba's own documentation directs developers to Qwen3.5-Omni for workflows requiring voice output. The context window is also structured rather than flat: 991,000 maximum input tokens, 131,000 maximum output tokens, and a maximum reasoning trace of 262,000 tokens.

Read more: Alibaba open-sources Qwen4 architecture preview that beats Claude Opus on agentic coding

The Pricing Comparison That Changes After December 31

The cost arithmetic is straightforward at the introductory rates. Qwen3.8-Omni-Flash charges $0.15 per million input tokens and $0.47 per million output tokens, with implicit cache hits billed at $0.016 per million. Gemini 3.8 Flash, released by Google on September 2, 2026, currently charges $0.75 per million input tokens and $3.75 per million output tokens under introductory pricing. That makes Qwen3.8-Omni-Flash 5× cheaper on input and approximately 8× cheaper on output at today's rates.

What the comparison omits is that Google's introductory rate has an explicit expiry. According to Google's Cloud Agent Platform pricing documentation, Gemini 3.8 Flash, Gemini 3.7 Flash, and Gemini 3.6 Flash all double in price on January 1, 2027 — input moving from $0.75 to $1.50 per million tokens, output from $3.75 to $7.50. Qwen3.8-Omni-Flash carries no comparable stated price expiry at launch. Teams building cost models today against the Gemini introductory rate should note that the comparison will shift significantly in three and a half months, with the sustainable gap between the two systems potentially widening to 10× on input tokens rather than 5×.

In concrete operational terms, Alibaba estimates audio input on Qwen3.8-Omni-Flash at under $0.01 per hour of audio, and 720p video with audio at one frame per second at approximately $0.20 per hour of content, not counting output tokens. The company claims a 98% reduction in audio cost per hour versus its own predecessor, Qwen3.5-Omni-Plus, and a 93% reduction in audio-visual cost. Both figures are company-reported and not independently verified.

The model reaches 113 audio input languages and dialects, accepts video files up to two hours and 2 GB by URL, and handles audio files up to three hours. It supports two-channel stereo and four-channel FOA spatial audio, which is relevant for meeting analysis that requires spatial localization of speakers. API compatibility spans both the DashScope and OpenAI protocols, the latter via Chat Completions and the Responses API.

What the Benchmark Table Says and What It Doesn't

Alibaba published a 29-evaluation benchmark table comparing Qwen3.8-Omni-Flash against Qwen3.5-Omni-Plus, Gemini 3.8 Flash, Seed 2.0 Lite, and Muse Spark 1.2. Across those evaluations, the company reports an average improvement of more than 25% over Qwen3.5-Omni-Plus, with the agent-specific gains reaching 36.5 points on WildClawBench-MM and 22.3 points on AgenticVBench. UniClawBench reaches 69.6. All figures are vendor-reported.

These numbers are meaningfully large if accurate, but several methodological constraints limit what the table can establish. First, every row was measured by Alibaba using its own evaluation infrastructure. Independent replication is not yet available. Second, BenchLM AI's review of the comparison notes that Alibaba ran different harnesses for different benchmarks — Claude Code for WildClawBench-MM and AgenticVBench, OpenClaw for UniClawBench — while applying provider-specific media settings for each competitor (Gemini 3.8 Flash at media_resolution=high, Seed 2.0 Lite at max_frame_tokens=384). The non-uniform settings mean the comparison is not reproducible by a third party without knowing the exact configuration Alibaba used for each competing system.

Third, on audio tasks where the company claims Qwen3.8-Omni-Flash exceeds Gemini 3.8 Flash, it is worth noting that the predecessor, Qwen3.5-Omni-Plus, already led Gemini 3.8 Flash on spoken-interaction benchmarks (92.9 vs. 92.3, per BenchLM's rendering of the evaluation), suggesting the audio advantage is not new to this generation. The novel claims are specifically in the agentic-audio-visual benchmarks.

For multi-speaker meeting transcription, Alibaba reports a dramatic improvement on AliMeeting — its own proprietary meeting evaluation dataset — with Diarization Error Rate falling from 88.11 to 3.35, and concatenated-word error rate improving from 89.61 to 17.18.

The practical recommendation from AiCybr's technical analysis is apt: "Independent benchmark evidence is still limited at release, making application-level evaluation the appropriate basis for production adoption."

Qwen-MM-Plugins: Bolting Omni-Modal Onto Any Agent Harness

The model ships alongside two open-source projects intended to reduce the integration friction for teams already running existing agent frameworks. Qwen-MM-Plugins is available on GitHub under Apache-2.0 and frames itself as "Make any agent harness multimodal-native."

The plugin architecture installs each capability as a Skill plus an optional MCP server. The guided installer supports Claude Code, CodeBuddy, Codex, OpenClaw, Qwen Code, and Gemini CLI — a deliberate breadth-first approach that does not require teams to migrate to a Qwen-native stack. Three core capabilities map directly to the launch's stated use cases: omni-memory builds a persistent audio-visual memory of a long video; omni-video2note converts a tutorial video into an illustrated PDF with timestamps; and omni-chatcut handles music video creation, movie commentary, and speaker-preserving video translation — the last of which requires the model to identify each speaker's vocal characteristics and re-dub translated dialogue to preserve the original voice profile.

Qwen-Live Harness, a second open-source release, enables real-time interaction through a connected camera and microphone — allowing agents to operate continuously on live video and audio rather than uploaded files. This positions the system for always-on monitoring, live meeting assistance, and interactive robotics-adjacent applications.

One practical limitation the README notes directly: most existing agent harnesses cannot feed audio to the model natively yet. Until harness-level audio support matures, audio is routed through the API rather than passed through the standard tool-calling interface. For video-only workflows this is not a constraint, but for combined audio-visual agents it adds a routing step.

Where Qwen3.8-Omni-Flash Sits in the Budget Multimodal Landscape

Prior to this launch, the budget tier for audio-video-capable APIs began at Gemini Flash pricing ($0.75/M input). Models with open weights — including DeepSeek V4 Flash Vision Exp (MIT license, $0.22/M off-peak) — offered vision-only capability without audio or native video support. The launch of Qwen3.8-Omni-Flash effectively creates a new price tier for the complete four-modality stack: below the Gemini Flash floor and without requiring the architecture workarounds that image-to-text bridges impose on audio content.

The tradeoff is the absence of open weights. Unlike DeepSeek V4 Flash Vision Exp, which can be self-hosted on a multi-GPU cluster for teams with on-premises infrastructure requirements, Qwen3.8-Omni-Flash is API-only at launch. Teams with data-jurisdiction requirements that preclude routing audio and video through an external provider — regulated industries, government contractors, organizations with EU residency requirements — cannot access the model's pricing advantage without accepting the external API dependency. The regional availability in six zones (Beijing, Singapore, Hong Kong, Tokyo, Frankfurt, and Virginia) gives developers in major markets a compliant routing option, but self-hosting is not yet on the table.

Alibaba has followed a pattern across the Qwen series of releasing proprietary API models before open-sourcing weight variants, and coverage of the model's launch notes that the company has said it plans to open-source the model following the pattern set by earlier Qwen releases. No timeline has been provided.

What Needs to Happen Next for This Market Position to Hold

Three developments will determine whether Qwen3.8-Omni-Flash's competitive position at $0.15/M tokens becomes a durable market fact or a temporary pricing event.

First, independent benchmark replication. The agent-perception claims are technically coherent, but no third-party lab has yet run the OmniVideoBench, WildClawBench-MM, or AgenticVBench comparisons against Gemini 3.8 Flash under controlled, reproducible conditions. Alibaba's prior benchmark releases have generally held up reasonably well under independent scrutiny, but the non-standard harness configurations in the current table make it harder to verify the specific competitive margins.

Second, the pricing stability question. Alibaba has not published a price-review or introductory-rate timeline for Qwen3.8-Omni-Flash. If the $0.15/M rate is itself a promotional figure tied to early-adoption acquisition, the cost calculus for teams making long-term infrastructure decisions could shift materially — just as Gemini Flash's doubling on January 1, 2027 will shift the Gemini calculus.

Third, open weights. The Qwen3.8-Flash-Next base is already available under the qwen-community-1.0 license, and the omni-modal variant adds a vision encoder and audio-text module on top. If Alibaba releases those weights, the competitive landscape for budget multimodal self-hosting changes significantly — particularly for teams that currently view DeepSeek V4 Flash Vision's 384-token image ceiling and absent audio support as limiting factors.

Until then, Qwen3.8-Omni-Flash represents a genuine step in making audio-video agent capability accessible at a price point that changes the unit economics of building with it — conditional on trusting company-reported benchmarks while waiting for the independent replication that determines how much of those numbers survive contact with an impartial harness.