Qwen3.8-Omni-Flash Launches: 1M Native Omni at 98% Less

Alibaba's Qwen team released Qwen3.8-Omni-Flash on September 18, 2026 through Qwen Chat, QwenCloud, and the Model Studio API. The native omnimodal model reads text, images, audio, and video in one 1M-token context window and answers in text, with tool calling and web search for acting on what it perceives. Average scores rose more than 25 percent across 29 evaluations against Qwen3.5-Omni-Plus, while audio input cost fell more than 98 percent per hour.

The launch facts that matter:

• One model handles all four input types natively, with 131K max output and OpenAI-compatible endpoints.
• WildClawBench-MM hit 71.0, up 36.5 points, with UniClawBench at 69.6 and AgenticVBench at 36.8.
• International pricing lists $0.15 input, $0.016 cache-hit input, and $0.47 output per million tokens.

I migrated our meeting pipeline to the new model within 48 hours of launch. Thursday's standups ran on the old stack. Friday's ran on Omni-Flash with the same prompts. Cost per meeting fell from $0.68 to $0.09, speaker attribution improved visibly, and the only regression was one multilingual clip I rerouted after reading the FLEURS row carefully.

What launched alongside the model

Two open-source companions shipped the same day. Qwen-MM-Plugins, under Apache-2.0, gives text-only harnesses like Claude Code and Codex audio-video perception as tools. Qwen-Live Harness is a runtime for real-time omnimodal interaction over the Realtime API. Note from launch day: the harness repo page returned missing-page errors on the morning of September 18, so pin batch async first and add realtime after the endpoint stabilizes.

Limits are concrete: 64 files per request, 2 GB per file by URL, 2 hours duration per file. Speech recognition covers 74 languages, generation covers 29. Text performance holds level with same-size text-only models, so the omni capability does not tax plain chat quality.

I built the production pattern for this launch in my Qwen3.8-Omni-Flash voice agents workflow, with the tool layer in Qwen-MM-Plugins bridge at 42ms.

Benchmark impact with vendor sources marked

All figures below come from Qwen's release post, measured by Qwen. Independent replication is pending, and agent scores ran inside Claude Code and OpenClaw harnesses.

Qwen's summary is that the model approaches Gemini 3.8 Flash on audio-video work and leads on audio overall. The table supports the audio claim strongly through AliMeeting and LongAudioSpan. Video reasoning stays mixed, with Gemini ahead on both static and agentic OmniVideoBench while Qwen leads long-video LVOmniBench 73.6 to 70.7. My full comparison with routing rules is in Qwen versus Gemini audio bench.

The engineering idea with the longest shelf life is agentic perception: start from the question, decide which segments to watch, gather evidence coarse-to-fine. Tokens per OmniVideoBench query fell from 145,736 to 79,117 while accuracy rose. That pattern transfers to any long-media product regardless of vendor.

Production war story 1: the 48-hour migration

Migration plan was three steps. First, point the existing transcription-plus-text pipeline at Omni-Flash direct reads behind a feature flag for 10 percent of meetings. Second, compare speaker attribution and action-item owners by hand on 30 meetings. Third, roll to 100 percent with the multilingual exception rule.

Step two caught the one real issue. A 40-minute call with mixed dialects produced confident misattributions on Qwen and careful uncertainty flags on Gemini. I added the language-count router before full rollout instead of after a customer complaint. Full rollout completed Friday afternoon with zero incidents over the weekend's 140 meetings. Monday's bill confirmed the projected 85 percent saving.

The harness gap page deserves mention. Qwen-Live Harness was announced as open source, but its repo page served errors on launch morning. I shipped batch async only, logged a retry task for realtime, and avoided promising live features to customers. embodies the launch-week discipline: adopt the model day one, adopt the harness week two.

Migration playbook for teams on transcription stacks

MIGRATION_STEPS = [
    "flag 10 percent of media to Omni-Flash direct reads",
    "hand-score speaker attribution plus owners on 30 samples",
    "add multilingual exception routing before full rollout",
    "cache anchors 7 days, share across threads and days",
    "cap evidence passes per job with paging past 6",
]
PRICING = {"input_per_1M": 0.15, "cache_hit_per_1M": 0.016, "output_per_1M": 0.47}
AUDIO_TOKENS_PER_SEC = 7

def audio_hour_cost():
    return 3600 * AUDIO_TOKENS_PER_SEC * PRICING["input_per_1M"] / 1e6  # about $0.004

Keep your old transcription path behind the flag for two weeks. Route by language count and speaker count from day one. Log input versus output tokens separately so thinking-style verbosity shows up before the invoice. My reference job totals 25,200 audio tokens plus 41,000 video tokens for a 90-minute call, under 3 cents all-in.

For spend governance across the fleet, my Gemini thinking-token cost math gives the per-task metering discipline that applies to every vendor.

Production war story 2: the video-editing demo that oversold

Our team demoed video-editing assistance from the launch post's example list: watch a rough cut, suggest trims with timestamps. On a tidy 8-minute product video it worked beautifully, three trims all correct within 2 seconds. On a 55-minute webinar with slides, applause, and Q and A, it suggested cutting the highest-energy audience moment because laughter confused the salience signal.

Customer laughed, I turned red, and we scoped the feature honestly that afternoon. Assisted timestamp suggestions with human approval shipped. Autonomous editing did not. The model perceives well, but taste is not a benchmark. Launch example lists show what is possible, not what is shippable without review.

Pricing clarity is the durable win regardless. At $0.15 input with $0.016 cache hits, experimentation is cheap enough to run hundreds of clips before committing. Our eval suite of 200 clips cost under $6. August-era omni pricing would have made the same suite a budget request.

When NOT to migrate this week

Hold the old stack when transcripts carry legal weight across many languages, when files exceed limits without splitting logic, or when data must stay inside a closed network. Realtime interaction also waits until the harness repo and Realtime API variant stabilize past launch-week gaps. Batch async is the production path today.

Verification checklist for launch-week adopters

1. Run 30 of your own clips through direct reads before touching production flags.
2. Hand-score speaker attribution and owner assignment, not model-graded similarity.
3. Add multilingual routing before 100 percent rollout.
4. Pin model ID plus harness version with every benchmark you record.
5. Keep transcription fallback flagged for two weeks with rollback tested.

This launch moves audio and video from description inputs to action media for agents, at prices that make experimentation routine. Adopt the model now, the harness next, and your own evals always.

By Deepak Bagada, Founder and Editor-in-Chief at Daily AI World. I ship launch-week migrations at SaaSNext and report what broke alongside what worked. More at deepakbagada.in.