Alibaba Launches Qwen3.8-Omni-Flash: Agentic Omni-Modal Model with Over 98% Price Cut

On September 18, Alibaba's Qwen team released Qwen3.8-Omni-Flash, billed as its first omni-modal model built around agentic capabilities: a single model takes text, image, audio, and video input, outputs text, and natively integrates understanding, reasoning, planning, tool calling, and result delivery. It is positioned directly against Gemini 3.8 Flash, with official-claimed capability jumps and a price cut of over 98% as the headline draws. Note: all confirmed information comes from official and relay posts; third-party evaluations have not yet appeared.

Confirmed

• Official accounts (@AlibabaQwen, @千问大模型) claim roughly 25% average gains over Qwen3.5-Omni with doubled Agent-related scores; @千问大模型 cites 30% over Qwen3.5-Omni-Plus
• Overlapping multi-speaker meeting transcription error rate drops from 88% to 3%; native support for up to 1 hour of continuous audio-video input; 1M-token context (per @eyishazyer: 991K max input, 131K output, 262K reasoning cap)
• Built on the Qwen3.8-Flash-Next architecture (per @johnseach)
• @testingcatalog relays that scores on WildClawBench-MM and UniClawBench are close to comparable Gemini models
• Audio-video Agent API pricing cut by over 98%; official performance/pricing pages and use cases published
• @solyarisoftware relays that the agentic video understanding mode saves 46% tokens while improving accuracy
• Accompanied by Qwen-Live-Harness and Qwen-MM-Plugins (per @karminski3); Qwen-MM-Plugins is Apache-2.0 open source, letting text-only agents like Claude Code, Codex, and Gemini CLI gain audio-video processing via a Skill plus optional MCP server (@eyishazyer)
• Officially positioned as a Gemini 3.8 Flash competitor with benchmark charts claiming multimodal wins
• Native output is text-only; media generation and multi-step tasks require function calls to external tools (@emmanuelvivier)

Unconfirmed

• Benchmark superiority over Gemini 3.8 Flash rests solely on official comparison materials with no third-party verification; relay posts (@solyarisoftware) caution that scores and availability need verification and question the "3.8" version numbering
• The 46% token-saving figure also comes from a relay post without direct official backing

Why it matters

• Cheap, natively agentic omni-modality dramatically lowers the barrier for real productivity scenarios like meeting transcription and long-video understanding; a 98% price cut could trigger a new price war among omni models
• Open-sourced Qwen-MM-Plugins extends audio-video capabilities to existing text-agent ecosystems like Claude Code and Codex, expanding Qwen's footprint in developer toolchains

2026-09-18 ~ 2026-09-19 · 15 related posts

• Alibaba Reportedly Drops Qwen 3.8 Omni Flash, Claims Multimodal Win Over Gemini 3.8 Flash — solyarisoftware · 2026-09-18
• Alibaba releases Qwen3.8-Omni-Flash, a cheap fast omni model for text, image, audio and video — gaganghotra_ · 2026-09-18
• Qwen3.8-Omni-Flash: meeting ASR errors cut from 88% to 3%, API prices down 98% — karminski3 · 2026-09-18
• [source] Qwen3.8-Omni-Flash Launches: Audio-Video Agent Model with 98%+ Price Cut — 千问大模型 · 2026-09-18
• [source] Alibaba launches Qwen3.8-Omni-Flash, its first omni-modal model built for agentic workflows — Alibaba_Qwen · 2026-09-18
• [source] Qwen shares pricing page for the new Qwen3.8-Omni-Flash omni-modal model — Alibaba_Qwen · 2026-09-18
• Alibaba Qwen shares performance, pricing and case details for its latest model — Alibaba_Qwen · 2026-09-18
• Alibaba ships Qwen3.8-Omni-Flash: 1M-token context, near Gemini 3.8 Flash on multimodal benchmarks — testingcatalog · 2026-09-18
• Alibaba launches Qwen3.8-Omni-Flash: omni-modal input, 1M-token context — emmanuelvivier · 2026-09-18
• Alibaba ships Qwen3.8-Omni-Flash: 1M-token omni-modal input with tool calling — emmanuelvivier · 2026-09-18
• Qwen3.8-Omni-Flash lands with agentic video understanding that cuts tokens ~46% — solyarisoftware · 2026-09-18
• Alibaba's Qwen3.8-Omni-Flash: 1M-context omni-modal, audio input 98% cheaper — eyishazyer · 2026-09-19
• Alibaba quietly ships Qwen-MM-Plugins: audio/video input for text-only agents — eyishazyer · 2026-09-19

2 near-duplicate retellings: karminski3 · johnseach