---

📊 Core Data Snapshot

• GPT-5.6 Luna Price Cut by 80% — Global large model price war begins
• 3 Domestic AI Video Tools Updated Same Day — Jimeng / MiniMax dual-line competition
• DeepSeek V4-Flash Context Window 1 Million — Developers' preferred high-cost-performance model
• 3 Top Companies Release Office AI Agents — B-end software track launches concentratedly

---

🔥 I. Today's Headlines

Today's three main threads in global AI software: Industrialization iteration of AI video, clustering launch of enterprise office agents, and global large model Token price-cut race.

Jimeng Seedance 2.5 Officially Released: 30-second Native Long Video + 3D Professional Plugins, Entering Film Industrialization

ByteDance launched Seedance 2.5 on July 31, supporting single-shot direct output of 30-second coherent scenes, with continuation generation up to 3 minutes; imports up to 50 reference materials, significantly optimizing character and scene consistency. Simultaneously adapted for Maya and Blender plugins, accompanied by updates to the Seedream 5.0 Pro image model; the new version increases credit consumption, raising the entry threshold for professional creators.

Industry Commentary | AI video tools are no longer limited to short-video entertainment, officially targeting advertising and film/TV B-end industrial production. ByteDance relies on underlying computing power to build a complete creation chain, but rising costs may lose lower-tier users; narrative consistency in long shots remains a core weakness.

DeepSeek V4-Flash Opens Public API Testing: Million Context, High Cost-Performance Capturing Developer Market

284B MoE total parameters, 13B active parameters, fixed lossless 1 million context window; code generation and agent task execution improved 6x compared to the preview version, natively compatible with OpenAI Responses API and Codex programming workflows, supports local private deployment, MIT weights free for commercial use, call pricing significantly lower than overseas closed-source models.

Industry Commentary | Domestic large model routes are clearly diverging. Kimi focuses on ultra-long document flagships, while DeepSeek anchors low-cost developer agent scenarios. This version directly impacts GPT lightweight versions and Claude entry-level models' overseas share.

Three Major Domestic Vendors Concentrate Updates on Office AI Agents: Doubao Enterprise Edition, 360 Nano Work, Meituan CatPaw Face Off

ByteDance's Doubao Enterprise Edition integrates with the full Feishu suite, adding data analysis, PPT generation, and process approval agents; 360 Nano Work launches isolated cloud virtual agent runtime environments; Meituan CatPaw covers 90,000 internal employees, building 30,000 local life service automation Agents. Analysys data shows domestic desktop office agent monthly visits exceed 60 million, with big players accounting for 93% of market traffic.

Industry Commentary | The AI office track shifts from internal trial-and-error to existing market competition. Patterns form in three sub-tracks: general office, vertical services, and code development. Survival space for independent small-to-medium office software vendors continues to narrow.

---

💻 II. Sector Highlights

• OpenAI Cuts Prices Across GPT-5.6 Series — Lightweight Luna cut by 80%, mid-range Terra down 20%, flagship Sol unchanged, adds high-speed Fast inference mode. Overseas leader initiates price defense; global Token price war intensifies.
• Kimi K3 Open Source Momentum Continues — Adds AgentEnv agent development sandbox suite; domestic weekly calls rank #1 globally for 13 consecutive weeks; People's Daily publishes industry commentary. Open source ecosystem consolidates long-text base; toolchain improvements raise migration costs.
• Alibaba Tongyi Qianwen Qwen3.6-Plus Iteration — Strengthens million-context programming capabilities; Bailian platform expands enterprise API quotas; simultaneously expands "Qianwen Office" enterprise beta. Connects model computing power + office scenario dual tracks.
• Google Releases Gemini 3.6 Flash Lightweight Model — Optimizes agent planning and code reasoning modules; introduces tiered low-price call packages for SMEs. Fills lightweight gap, forming a complete flagship + lightweight matrix.
• MiniMax Releases H3 Full-Modal Video Model — Natively outputs 2K/15-second stereo video; priced at only 1/3 of overseas equivalents; previews open-sourcing weights soon. Differentiates from Jimeng, capturing lower-tier short-video track with low-price open source.
• Tencent WorkBuddy Continuous Iteration — Deeply bound to Tencent Docs ecosystem; iterates real-time human-machine co-creation and cross-format batch processing; long-term key tested product by ShuJie Workshop. Stabilizes B-end base relying on native traffic from office software.
• ComfyUI 0.29.0 Update — Adds video training nodes; optimizes memory scheduling; significantly reduces VRAM usage for long video generation. Open source tools continuously improve underlying infrastructure, promoting popularization of personal AI video creation.
• NetEase Youdao Integrates DeepSeek V4-Flash Across All Products — Dictionary, essay correction, and coding tutoring uniformly switch to new model; upgrades document summarization, logical reasoning, and code error correction capabilities. Vertical education software batch-integrates domestic high-cost-performance models.
• Huawei Ascend Atlas 950 Super Node Lands — 8192-card interconnection completes live verification; natively compatible with DeepSeek, Jimeng, and Qianwen series private deployments. Hardware-software integrated support system forms; enterprises prioritize Ascend solutions for private deployment.
• Anthropic Discloses Claude Opus 5 Iteration Plan — Surpasses predecessor on multiple benchmarks; price halved; receives $5 billion computing power support from AMD to counter NVIDIA + OpenAI alliance. Opposition between two major overseas computing-model camps intensifies.

---

⚡ III. News Briefs

AI Large Models

• OpenAI internally tests multi-agent collaboration model Astra, supporting decomposition of complex long-cycle projects (official test disclosure)
• OpenRouter data: Total call Token volume of domestic models far exceeds US vendors (OpenRouter statistics)
• Moonshot AI completes $3.5 billion Series F funding, advancing HK Pre-IPO as fast as 6 months (financial media)

AI Software & Office Agents

• Meituan CatPaw opens paid enterprise version to local life merchants, providing dedicated Agent hosting services (Meituan official)
• Meta AI launches paid API pricing system, benchmarking OpenAI's commercialization model (overseas tech media)
• KeDa ZhiLing completes three rounds of financing, focusing on team-level AI code R&D platforms (36Kr)

Video Generation AI Tools

• Jimeng opens separate subscription for 3D plugin paid membership, no need for full premium membership (Jimeng backend announcement)
• Midjourney V7 minor update to character consistency algorithm, lowers base unit price for image generation (Midjourney Discord)
• Runway Gen4 optimizes long-shot smoothness, adds automatic film storyboard generation feature (Runway website)

---