01MODEL RELEASES3 stories
DeepSeek 9/14 flagship retirement-style switch: V4 Pro fully routed to V4.1 Flash
DeepSeek官方公告+行情Confidence HighProgress · Day 2
Official API logs confirm that from 9/14, deepseek-v4-pro is fully routed to V4.1 Flash and priced at Flash rates (official model card Terminal-Bench 4.0 = 31.2); the official announcement discloses the new architecture's KV cache has HBM demand only 1/4 of the old version, and SSD 1/8.
Take · Memory chain weakens in response (SK Hynix US shares -5.2%, Micron -4.9%, Seoul down another 3%+ on Friday) — algorithmic efficiency for the first time strikes back at the HBM shortage price-hike narrative; note the split in evaluation methodology, AA's independent methodology is clearly lower than the official model card, citations must include harness.
ElevenLabs官方视频Confidence High
ElevenLabs official video releases Music v2.5 (Introducing Music v2.5).
Take · Music generation model iteration accelerates; audio generation becomes a stable release surface on official channels.
Original ↗OpenAI官方视频Confidence High
OpenAI announced in an official video that GPT-Live-1 is now in the API.
Take · Voice capabilities are moving from an in-product feature to a programmable interface, allowing voice Agent call chains to be independently integrated and priced.
Original ↗02PRODUCT3 stories
OpenAI pauses new subscriptions and upgrades for $200/month Pro 20X
OpenAIX 官宣 / PCMag / 多家外媒Confidence High
Core product lead Sottiaux announced on X on 9/10: due to 'unprecedented' demand for GPT-6 Astra, new subscriptions and upgrades for $200/month ChatGPT Pro 20X are paused (existing subscriptions unaffected). Deep version note: a third-party single-source estimate suggests Plus/Pro 5X enters negative gross margin territory once usage exceeds about 11.4% (low confidence, for lead purposes only).
Take · The flagship subscription tier is being overwhelmed by its own demand, corroborated by the Token ledger (median researcher daily inference cost >$600)—the compute bottleneck has moved from rumor to hard fact, and workflows heavily dependent on this tier need downgrade contingency plans.
Ant Group sets the tone for the 'agent commerce moment': Alipay establishes an annual 10 million yuan 'Agent Emergence Award'.
蚂蚁 / 支付宝第一财经 / 新华财经Confidence High
Ant Group CEO Han Xinyi announced at the Bund Summit on 9/11 that 'the ChatGPT moment for agent commerce has arrived,' flagging the bottleneck: 'Users want to use it, but merchants don't have good agent services'; Alipay is establishing an annual 10 million yuan 'Agent Emergence Award' (single award up to 1 million yuan).
Take · Supply-side 'Agent-callable transformation' (AEO) has become a clear white space—retrofitting merchant services so Agents can understand and call them is the SEO-level opportunity right now.
百度智能云官方视频（B站）Confidence High
Baidu AI Cloud released a series of official videos on 9/12: Agent enters real life (from 'My Companion' to 'Our Family Companion'), Xiaodu Tiantian Guimi Machine Ultra (move freely, AI on the go), Xiaodu Smart Speaker Pro Max (emotional, perceptive), and previewed the 9/16 2026 Smart Economy Forum.
Take · Home-scenario Agents are moving from single-point hardware to a “mobile screen + speaker + companionship” product matrix; the 9/16 forum is the next watchpoint for industry agents.
Original ↗03INDUSTRY8 stories
AI agents weaponized at scale for the first time: PaperCut vulnerabilities breached 395 organizations in 48 countries
GreyNoise / Blackpoint双源安全厂商Confidence High
Two sources confirm: attackers exploited two PaperCut vulnerabilities (CVE-2026-81578/82078), letting AI agents based on OpenAI Codex (harness) and DeepSeek (model) automatically complete vulnerability development and execution, breaching 395 organizations in 48 countries within hours; a U.S. high school gained domain admin access in 7 minutes (only 12 of 440 instances had full domain admin; citations must include the stated scope).
Take · 'Agent weaponization' has moved from simulation to real incidents: Agents' API keys and automation permissions are the new attack surface; security audits and permission controls are becoming the next must-have category.
Pentagon plans to provide about a $5 billion loan to cloud startup Fluidstack
五角大楼 / Fluidstack华尔街日报 / 路透社Confidence High
WSJ first reported on 9/10, Reuters followed and confirmed: The U.S. Department of Defense, through the Office of Strategic Capital (OSC), is in talks to provide about a $5 billion loan to cloud computing startup Fluidstack for data center component supply chain (not new construction); still under negotiation and not finalized.
Take · If completed, it would be the U.S. government's most direct financing move for private AI infrastructure—AI infrastructure moving from 'VC logic' to 'state credit logic,' forming a mirror image of DeepSeek's architecture-level demand reduction.
Moonshot AI's August ARR reached $1 billion, targeting $2 billion by year-end
月之暗面BloombergConfidence HighProgress · Day 1
Bloomberg 9/11: Driven by the July launch of Kimi K3, Moonshot AI's August annualized recurring revenue reached $1 billion (June $300 million, up more than 2.3x in 5 weeks), with plans to reach $2 billion by year-end; it is advancing financing at a $50 billion pre-money valuation and may list in Hong Kong within the year.
Take · Adds revenue-side validation to the 'A+H dual-listing exploration' reported on 9/11; P/ARR remains the highest tier in the open-source camp, and the next checkpoint is whether month-over-month growth holds up.
Two major IPOs together: Anthropic moves up to October; OpenAI files confidentially.
Anthropic / OpenAI路透 / FT 转述多源Confidence Mid
Reuters confirmed on 9/12: Anthropic's IPO window moves up to a mid-October roadshow and completion before the midterm elections, raising up to 1000 亿 and valued at about 2 万亿, with NVIDIA planning a 100 亿 cornerstone investment; on the same day, multiple sources confirmed OpenAI confidentially filed its S-1 on 6/8.
Take · For the first time, the safety narrative and capital narrative are juxtaposed and offset each other, resetting the global AI valuation anchor in Q4; the cost/safety disclosures in the prospectus are more worth reading page by page than the launch event (OpenAI's filing date is subject to official disclosure).
First bank agent foundation order settled: 6489.61 万元 winning bid
浙商银行 / 北京可利邦财联社 / 新华财经Confidence HighProgress · Day 4
The Zheshang Bank agent centralized procurement reported on 9/8 is settled: 'agent foundational software and hardware' won by Beijing Kelibang at 6489.61 万元 (a new single-order high for financial institution agents this year), plus Ant-affiliated 946 万 wealth/supply chain orders totaling about 7436 万.
Take · Bank procurement shifts from 'buying models' to 'buying the full foundation' (GPU compute pool + knowledge engineering + model scheduling + tool calling), with demand mainly from urban and rural commercial banks—a dense period for financial Agent tenders begins.
August CPI released: software component +25.4%, AI costs enter inflation gauge for the first time
市场（BLS / CME）BLS / CME FedWatch / 财经媒体Confidence HighProgress · Day 2
US August CPI released on 9/11: headline 3.4% in line with expectations, core MoM +0.3% above expectations (largest monthly gain since April); probability of a 25bp FOMC rate hike on 9/16 rose late in session to about 87%-90% (CME snapshots at different times); CPI component 'consumer software and accessories' +25.4% YoY, a record high.
Take · AI costs formally enter the inflation gauge, tightening the rate-sensitive window for AI stocks; US stocks rose on the day as the 'bad news landed', ending a four-day losing streak.
MIIT 'AI+Software' special action: cover 20,000 above-scale software enterprises by 2028
工信部工信部发布 / 新华社Confidence High
Implementation Plan for the 'AI+Software' Special Action (信发〔2026〕209 号, issued 9/2, publicly released 9/11): by 2028, cover 20,000 above-scale software enterprises, 100 technical upgrades, 100 agent benchmarks, 5 open-source projects, with supporting computing power vouchers; first systematic deployment of security review for AI-generated code and agent identity identification/trusted interconnection/behavior control.
Take · Compliance reach extends into the 'AI coding' supply chain: code security review and agent identity become the next must-answer items in bids, and a new paid category (note the date gap between issuance and public release).
California bans AI companion toys + ASI bill sets penalties
加州 / 美国参议院加州官方 / Sanders fact sheetConfidence High
California signs SB 867: first in the U.S. to ban AI companion chat features in toys for children under 16 (for 5 years, no disclosure/certification exemption path; the feature itself is a violation), affecting the entire toymaker-AI supplier-retailer chain; meanwhile Sanders/Casar unveil the framework for the 'Ban Artificial Superintelligence Act': corporate dissolution-style 'death penalty' + up to 20 years' imprisonment for individuals (proposed 9/3, bipartisan briefing 9/16; still a proposal, not law in effect).
Take · For the first time, 'software behavior' is regulated as physical product safety; makers of children's/companion AI products must immediately self-check against California's stance—the feature itself is a violation, with no 'disclose and you're fine' backdoor.
04RESEARCH2 stories
Evaluation paradigm shifts to harness: HarnessDev benchmark + 方升-Code 2.0
字节 Seed / 信通院encorp.ai 转述 / 腾讯新闻Confidence Mid
ByteDance Seed and multiple institutions release HarnessDev benchmark: same weights, different execution environment, score gap nearly 15 points (GPT-5 scores 35.2% under Terminus 2, rises to 49.6% with Codex CLI); CAICT releases '方升-Code 2.0' (5000+ real project data, adds engineering quality/stability dimensions).
Take · Evaluation targets shift from “answers” to “runnable systems”—the harness itself becomes the test, and citing model scores must include the execution environment.
Cohere官方视频Confidence Mid
Cohere official post (Lukas Kuhn): LeVJEPA—Efficient & Scalable Video Pretraining without the Heuristics.
Take · An official technical signal for the “fewer assumptions, scalable” route in video pretraining.
Original ↗05INSIGHTS5 stories
Today’s main thread: the day AI agents are weaponized is when state credit enters.
前瞻深度版研判Confidence Mid
PaperCut agent weaponization (already occurred) × Pentagon $5 billion loan (state credit logic) × Ant “agent commerce moment” on the same day across three threads: the theme shifts from capability narrative to three hard constraints—safety, capital, and compliance.
Take · Deep-dive grey-area judgment: the certainty narrative is “compute supply-demand gap and price hikes, DeepSeek architecture-level cost reduction, Agent safety regulation tightening (synchronized domestically and internationally)”; IPO timing and V4.1’s true capability ceiling remain to be verified.
Seize AEO: make enterprises 'agent-callable' service providers
行动深度版建议Confidence Mid
Start with merchants that have online APIs (OTA/financial products/local life), and build the trio of interface standardization, structured data, and Agent trust credentials; path: free 'Agent reachability' diagnostics for 3-5 companies → deliver transformation and accumulate SOPs and toolchain → platform-based certification services, benchmarked against early SEO.
Take · Ant setting the tone is official signaling; shutdown signal = a big tech firm launches an official AEO standard. AEO naturally fits per-call/per-transaction billing and synergizes with the 'performance-based billing' recommendation.
Be a subcontractor for 'tool calling + knowledge engineering' in banks' agent foundation
行动深度版建议Confidence Mid
Use Zheshang Bank's 6489.61 万 winning bid as a tender template, and target the most understaffed modules in the foundation: in-bank knowledge base RAG governance, tool-calling gateway, and agent scheduling/monitoring; path: target 10 joint-stock/city commercial banks → after 2-3 deals, productize the tool gateway → embed into MIIT's 'AI+Software' special application to obtain computing power vouchers.
Take · Banks have just begun shifting from buying models to buying foundations, with 2026-2027 a dense tender period; shutdown signal = leading big banks become monopolized and entrenched by the four major integrators. Bundle Agent security audits (the PaperCut lesson) into every delivery—security is a mandatory question for banks.
Bet on 'performance-based billing' infrastructure: workload metering and result attribution tools
行动深度版建议Confidence Mid
Be the 'electric meter' of the agent era—API call metering, result attribution, and multi-party split settlement middleware; path: single-scenario MVP (Token split settlement) → integrate with 1-2 agent platforms for validation → strive to become a drafter of industry metering standards.
Take · Shanghai Advanced Institute of Finance × Ant Research Institute report systematically endorses “pay-for-performance”; demand hardens for workload metering, outcome attribution, and multi-party settlement tools; it is the shared foundation for both AEO and the banking base.
Verification note: 15 verified today, 6 accepted / 5 corrected / 2 deferred / 2 downgraded
提示深度版核验Confidence Mid
Key points: rate-hike probability uniformly stated as “about 87%-90%” (CME snapshots at different times); the MIIT plan was issued on 9/2 and publicly released on 9/11, not issued on 9/11; Ant has an annual 1000 万元 “Agent Emergence Award” fund (not a “1000 万 award”); penalties in the “Banning ASI Act” come from the 9/3 proposal; DeepSeek evaluations follow the official model card (AA 39.55 / LiveBench 77.3 not independently verified; not accepted).
Take · Downgraded/removed: the Computing Power Conference’s “demand +417% vs supply +128%” was old Q1 data re-cited, downgraded to background; Discovery Loop’s 500 亿 valuation has only Business Insider as a single source (asking price, not a completed deal), deferred; Qwen iris payment glasses downgraded due to “may become” promotional framing; OpenAI cutting off Cursor (reported 9/1) has no increment, not repeated.
Ecosystem Pulseself-updating
News299Skills72OpenHub44Reddit252026-09-02 → 2026-09-10 Star growth：obra/superpowers +90,630★ · mattpocock/skills +82,178★ · multica-ai/andrej-karpathy-skills +67,613★ · anthropics/skills +56,067★ · Shubhamsaboo/awesome-llm-apps +43,732★
✦Action Items3
1. Capture AEO: make enterprises “agent-callable” service providersStart with merchants that have online APIs (OTA/financial products/local services) and deliver the three-piece set of interface standardization, structured data, and Agent trust credentials; first provide free “Agent reachability” diagnostics for 3-5 companies, use delivery transformations to consolidate SOPs and a toolchain, then offer platform-based certification services.
window Ant setting the tone marks the start of official signaling; shutdown signal = a big tech company launches an official AEO standard 2. Subcontractor for 'tool calling + knowledge engineering' in the bank agent foundation.Use Zheshang Bank's 6489.61 万 winning bid as a proposal template; target the three most understaffed modules—in-bank knowledge base RAG governance, tool-calling gateway, and agent scheduling/monitoring; after landing 10 joint-stock/city commercial banks, productize on the basis of 2-3 orders, and embed into MIIT's 'AI+Software' special application to obtain compute vouchers.
window 2026-2027 is a dense period for bank tenders; shutdown signal = leading major banks monopolized and locked in by the Big Four integrators. 3. Bet on 'performance-based billing' infrastructure: workload metering and outcome attribution tools.Build the 'electricity meter' of the agent era—API call metering, outcome attribution, and multi-party split settlement middleware; start with a single-scenario MVP (Token revenue-share settlement), integrate with 1-2 agent platforms for validation, then seek to become a drafter of industry metering standards.
window Metering tool first movers are scarce; shutdown signal = cloud giants open built-in settlement capabilities for free.