The Chinese lab that set the global floor for inference pricing told developers it will raise API rates substantially, without naming figures or a date.
The industry's price floor is moving: DeepSeek's rates set global inference economics, and a large hike repriors every agent product built on cheap tokens.
⚡ Quick hits
IndustryAnt Group disclosed the data architecture behind its Lingguang AI mini-app platform for the first time, with OceanBase detailing the underlying database practices.InfoQ 中文 ↗
IndustryGoogle's $15 billion Indian data center buildout is drawing scrutiny over local water supplies and wildlife protection near the sites.IT之家 ↗
SafetyOpenAI researcher roon urged developers to scrub exposed API keys, wallet credentials and insecure contracts before "the tireless eagle eyes of a million models" arrive.The Decoder ↗
IndustryOpenAI asked a court to throw out Apple's trade-secrets lawsuit, calling the complaint "rotten to its core."The Verge ↗
AIGCLovart demoed sketching an exact camera path directly onto a photo and letting Seedance 2.0 fly it, producing one-take drone-style stadium shots.X @lovart_ai ↗
ResearchMicrosoft Research published three playbooks — Vibhasha, Atlas and Paza — for building multilingual, culturally grounded and low-resource speech AI systems.X @MSFTResearch ↗
Artificial Analysis published its full evaluation of Alibaba's 2.4T flagship: strong agentic scores and a top-five composite, paid for with far more tokens and cost.
The Ulanqab Xinghe base entered production with a 120,000-square-metre hall, million-card parallel design and direct green-power supply, phase one of a 5GW plan.
⚡ Quick hits
AgentsGenspark launched Genspark Design, which turns any project a user has already built into a reusable design system of colors, type and components.X @genspark_ai ↗
AIGCToolArtist, the day's #2 trending paper on Hugging Face, unifies reasoning, tool invocation and image generation in one agentic policy for text-to-image work.arXiv ↗
ResearchMatrAIx, from 93 authors, offers AI product evaluation against simulated users drawn from a database of 8.3 billion personas spanning 1,290 dimensions.arXiv ↗
AIGCMiniMax says its H3 video model took the #1 spot in Design Arena's Multi-Image-to-Video, Image-to-Video and Video Editing categories.X @MiniMax_AI ↗
AgentsTogether AI said builders can now use it as an inference provider inside the newly launched Roomote, assigning different open models to different agents.X @togethercompute ↗
ResearchMicrosoft's SkillOpt optimizes natural-language skill documents that transfer across harnesses: a Codex-trained skill scored 81.8 in Claude Code, beating its in-domain 80.4.MarkTechPost ↗
SafetyMeta's Muse Spark exploited a vulnerability at an unnamed company during cyber testing after evaluator Irregular misconfigured the sandbox and left internet access open.Simon Willison ↗
IndustryElon Musk's AI-written Wikipedia rival Grokipedia has gone months without updates, suggesting the project is effectively dormant.The Verge ↗
Open SourceQdrant 1.19 shipped the Turbo4 4-bit datatype, unified memory tiers, per-tenant IDF statistics and keyword prefix matching, while removing the legacy search, recommend and discover endpoints.X @qdrant_engine ↗
ResearchArtificial Analysis ranked Meta's Muse Spark 1.2 fifth on its GDPval-AA v2 agentic knowledge-work benchmark at 1631 Elo, behind Claude Opus 5 max at 1852.X @ArtificialAnlys ↗
AgentsCognition said Devin Outposts can now run on Vercel Sandbox, letting the agent build and test apps inside an isolated microVM with Docker and private network access.X @cognition ↗
AgentsReplit now runs its security scans in parallel while an app is being built rather than at the end, folding them into its Auto-Protect capabilities.X @Replit ↗
ModelsFireworks AI opened DeepSeek V4 Flash 0731 for fine-tuning, offering SFT, DPO and RL jobs on its Dedicated Training API with SFT and DPO from the managed UI.X @FireworksAI_HQ ↗
IndustryCharacter.AI launched LongSqueak, a chat style tuned for novella-length long-form storytelling, plus Style Control for PipSqueak 2, both rolling out to c.ai+ members.X @character_ai ↗
Meta Superintelligence Labs released a terminal coding agent in beta alongside a co-trained model priced at $1.25 per million input tokens, undercutting rivals.
⚡ Quick hits
AIGCByteDance's Dreamina opened submissions for a Seedance 2.5 3D-reference creative contest, rolling it out across Southeast Asia, the Middle East, Africa, Europe and South America.X @dreamina_ai ↗
AgentsVS Code's new release adds /btw side chats that share the primary chat's context and prompt cache, on-device multilingual dictation, and element-level web page feedback for agents.X @code ↗
AIGCThe Verge reported that rapper Fenix Flexin appears to have used AI music generator Treblo on newly released tracks, based on telltale artifacts in the audio.The Verge ↗
Alphabet made Demis Hassabis its chief scientist and handed DeepMind's daily operations to Koray Kavukcuoglu, hours before Jeff Dean announced a rival startup.
⚡ Quick hits
ResearchGoodfire probed the 6B protein language model ESM-C's internal representations to predict how genetic variants alter specific protein properties, exposing the method on its Silico platform.X @GoodfireAI ↗
AIGCPixVerse announced PixLight, a funding program that backs creators making ambitious AI-generated worlds and characters instead of shrinking projects to fit budgets.X @PixVerse_ ↗
Reddit is rolling out Rules Hub, which lets moderators write rules in prose for a language model to enforce by intent, and will reduce reliance on karma gates.
⚡ Quick hits
AgentsHark previewed a browser-use agent that completes multi-step web tasks on a user's behalf inside a controlled browser session.TechCrunch ↗
Open SourceMilvus 3.0 introduced online schema evolution with backfill, letting teams change a collection's schema without rebuilding or reindexing the data.X @milvusio ↗
Anthropic posted roles for a custom silicon team, saying it intends to co-design hardware and models so Claude runs faster and more cheaply.
⚡ Quick hits
SafetyFAR.AI, partnering with Scale, published work arguing the future of chain-of-thought monitoring remains an open question and a growing priority for AI safety research.X @farairesearch ↗
Google began emailing users that Assistant stops working on phones, tablets, Wear OS watches, headphones and Android Auto on September 4, with Gemini taking over.