01MODEL RELEASES4 stories
StepFun StepAudio 3 launches five speech models at once, tops AA speech leaderboard
阶跃星辰多源Confidence HighNewcomer
Five speech models released together, topping multiple Artificial Analysis speech leaderboards: Realtime dialogue dynamics 98.9% first, ASR WER 1.7% tied for first; Realtime supports native full-duplex + asynchronous Tool Call.
Take · The speech layer upgrades from 'ears and mouth' to 'reasoning while listening'—native full-duplex plus asynchronous tool calling means a speech Agent can call the backend directly during a call, the first hard metric for the day's 'cost inflection point in speech and Agent scenarios'.
Unisound U2-Flash: 266B total parameters / ~10B activated, Agent Token down 20-30%
云知声港交所公告（一手）Confidence High
Small-activation MoE route: 266B total parameters, ~10B activated; Agent iteration steps and Token consumption both down 20-30%, DeepSWE v1.1 score 64.6.
Take · The other half, isomorphic to StepAudio 3: the speech side drives down real-time costs, the model side drives down Agent iteration costs; together they form the economic basis for 'small, fast, cheap' to be quote-ready in Q4.
Zidong Taichu 5.0-9B open-sourced: No. 1 among models under 10B on 8 of 9 spatial understanding benchmarks
紫东太初多源Confidence High
Open-source small model for spatial understanding ranks first in 8 of 9 benchmarks under 10B parameters, with public weights.
Take · Open-source competition continues to sink into the 'vertical capability + small parameters' layer, complementing today's 小红书 Iris and 书生-S2 search/research lines: a privatizable spatial understanding foundation emerges, giving embodied/inspection projects another domestic option.
LMArena 9/15 refresh: closed-source leaders break 1500 Elo for first time, Kimi K3 touches 1500
LMArena榜单快照多源Confidence MidMethodology downgrade
Closed-source leaders break 1500 Elo for the first time (Anthropic occupies the top three); on the open-source side, Kimi K3 touches 1500—the preference gap between open and closed source enters the single-digit noise band.
Take · Today's reporting ledger has downgraded 'three models break 1500 for first time' to 'closed-source leaders break 1500 for first time' (aggregator-site methodology noise); when citing leaderboard scores, always include snapshot date and methodology—this is part of the day's evaluation trust crisis.
02PRODUCT4 stories
Waymo's tale of two cities: Las Vegas opens fully driverless to the public, teams up with Japan targeting 2027 Tokyo commercial service
WaymoReutersConfidence HighProgress
Las Vegas opens fully driverless service to the public for the first time (6th-gen Driver / Zeekr platform); teams up with Japan's GO and Nihon Kotsu, targeting Japan's first fully driverless commercial service in Tokyo in 2027.
Take · Dual tracks—U.S. interstate replication and East Asia overseas expansion—are accelerating simultaneously, showing the Robotaxi bottleneck has shifted from technical validation to 'city access + partner licenses,' which also explains the same-day surge in overseas orders from China's intelligent-driving supply chain (Hesai/Joyson).
NVIDIA DSX AI factory's first production validation: throughput +24% / per-watt +23%
英伟达官方博客Confidence High
Lambda B200 cluster deploys 19 nodes at the same power, throughput +24%, per-watt performance +23%; 'AI factory × grid' demand-response real-world run completed for the first time (automatic load reduction 4MW→3MW); AI Infra Summit drew 8000 attendees (3500 last year).
Take · Writing output per watt and grid demand response into the official report card amounts to acknowledging that power and grid-connection capacity have become hard constraints on compute expansion—two sides of the same coin as the EU adding advanced computing ICs to its dual-use list.
Salesforce × Google Cloud three-layer rollout: Hyperforce moves to GCP, MCP interoperability, UCP enters search
Salesforce × Google Cloud官方新闻稿Confidence High
Hyperforce handles production traffic migration to GCP (North America GA in November); accesses Gemini Enterprise via MCP headless-architecture interoperability; UCP protocol brings Commerce Cloud into Google Search AI Mode.
Take · Ecosystem power shifts down to the orchestration/interoperability layer: MCP and UCP are simultaneously treated by two giants as access surfaces, meaning the protocol layer is replacing the application layer as the bargaining point—this is also why MCP proxy appeared in that day's government and enterprise procurement list.
Agility Digit 5: fenceless human-robot collaboration architecture, payload +40%, 90-minute runtime
Agility Robotics官方 + Unite.AIConfidence HighScope has been defined.
The new architecture uses AI human monitoring + an independent safety controller to enable fenceless collaboration; payload up 40%, runtime 90 minutes. Note: the 'over $300 million order' is a restatement from May SPAC materials, not new.
Take · The daily ledger has removed the order figure from incremental items (restatement item), keeping only hardware parameters: credible metrics for humanoid robots still center on 'safety architecture + payload + runtime'; commercial figures await newly signed primary materials.
03INDUSTRY11 stories
OpenAI reportedly seeking $1.2 trillion pre-IPO financing, +40% vs March valuation
OpenAIFT 9/15（一手已核）Confidence High
FT: OpenAI in talks with large investors for pre-IPO private financing, targeting a $1.2 trillion valuation (+40% vs $852 billion post-money in March), investor-initiated and at an early stage; 数见智 verified to FT first-hand, not a The Information restatement.
Take · Not contradictory to Altman's 'no IPO in 2026'—it is precisely the scramble-for-stakes logic against a delayed listing backdrop. Trillion-scale valuation and 92% rate-hike probability in the same frame: during liquidity tightening, leading players are still draining the whole market's liquidity.
FOMC eve: tightening pricing maxed out at 92%
美联储 / 宏观财联社 / 新华社Confidence HighProgress · Day 2
9/15 U.S. session: 10Y hit 5.04% intraday (revised upward to 'highest since 2007'), CME prices 92% odds of a 25bp hike, 17 of 20 investment banks bet on it; WTI settled at 105.83 (108.59 was misrecorded), diesel futures hit a record high; tonight's retail sales (expected +0.9%) and tomorrow morning's decision land on the same day.
Take · Whether it is bad news exhausted or a second shock will be set by tonight's retail sales + tomorrow morning's dot plot; the deep-dive edition lists 'liquidity tightening hitting AI valuations' as the day's No. 1 gray rhino (75-85%).
Factory completes $200M financing, valuation 5 billion (3.3x in 7 months)
FactoryReutersConfidence High
April: 1.5 billion → 5 billion valuation, about 3.3x in 7 months; Blackstone, Khosla, Sequoia, Insight participated.
Take · After Cognition’s 48 billion, PE giants are pricing coding agents as infrastructure—“Agents are software assets” has become a primary-market consensus, and the valuation anchor is shifting from revenue multiples to infrastructure metrics.
Profound’s $180 million Series D at a $1.8 billion valuation: the AEO/GEO sector’s first billion-level round
Profound官方通稿Confidence High
Co-led by Sequoia + Kleiner Perkins; serves 1,000+ enterprise brands (about 1/3 of the Fortune 100); less than 7 months after its previous $96 million Series C.
Take · “Getting products recommended by AI” has been confirmed by capital as an independent track (AEO/GEO), while there is no leading player in China yet—the deep-dive version lists “AEO / multi-model routing services” as one of its Top3 action recommendations, with a window of about 6-12 months.
DeepSeek plans to appoint its first CFO (plans to hire, unconfirmed)
DeepSeekReuters / 36氪Confidence MidProgress
Reuters independently reported on 9/14 that it plans to hire Gaorong partner Yan Wentao (born 1991; invested in MiniMax / Zhipu) as its first CFO.
Take · Another institutional signal for the STAR Market path—installing a CFO is a preliminary step in the capitalization process; the wording is “plans to hire,” and neither the individual nor the company has confirmed it, so it cannot be cited as a fait accompli.
Government and enterprise AI procurement moves down to the 'connection layer': MCP agents enter procurement lists for the first time
中原银行 / 国寿养老官方公示Confidence HighNewcomer
Zhongyuan Bank's 87 万 'enterprise-grade AI gateway' winning bid candidate announcement; for the first time, the procurement list includes MCP proxy / AI Agent proxy; Tongchuang China Life Pension's 86 万 agent cluster management and control platform, SAIC Asset's 320 万, Shantou Customs' 96.74 万, and other government/enterprise AI orders in the 50-320 万 range landed in quick succession.
Take · AI gateways are replicating the independent category path of API gateways from back then; with a 50-100 万 price point and 30 天 delivery, they are a subcategory that small and medium software vendors can standardize; the in-depth version lists it as a 'best bet this week' action recommendation.
Chinese smart-driving supply chain's two overseas orders: Hesai's USD 40 million exclusive order + Joyson's ~RMB 15 billion design win
禾赛 / 均胜电子证券时报Confidence High
Hesai wins an exclusive lidar order worth over USD 40 million from a leading US Robotaxi company (sole supplier, delivery by end-2026); Joyson Electronics newly wins intelligent designations from two leading OEMs, worth about RMB 15 billion over the full lifecycle.
Take · Complementing Waymo's two-city expansion: the direct beneficiaries of the overseas Robotaxi ramp are China's perception and cockpit supply chains; overseas orders have moved from samples to exclusive + design wins.
Regulatory three-way split: Framework 3.0 released, U.S. federal window closes, EU controls expand
中国网安标委 / 美国 / 欧盟人民日报 / The Hill / 欧委会一手Confidence HighProgress · Day 2
Cybersecurity Standards Committee's 'AI Safety Governance Framework 3.0' released on 9/14 at Cybersecurity Week, endorsed by MFA on 9/15; Trump on Truth Social explicitly opposes federal AI legislation + House recessed for 7 weeks; European Commission dual-use list adds advanced-computing ICs and EUV masks / ALD deposition / single-wafer cleaning equipment (after 2 months of review, pending publication in the Official Journal).
Take · Three jurisdictions gave three answers in the same week: standards iteration tightening (China), legislative vacuum (U.S.), export-control expansion (EU); the direct implication for companies is to prepare compliance baselines to the strictest interpretation while seizing procurement windows to the loosest—'institutional vacuum + cash-flow pricing' is the day's main logic.
AI workforce structural polarization: aggregate not collapsing, structure shifting sharply
韩国央行 / Conference Board / 脉脉BOK / PR Newswire / 界面Confidence HighThree-source corroboration
Bank of Korea's first regional report: 80.8% of the 247,000 new jobs in high-exposure occupations are concentrated in the capital region; high-exposure employment for youth in their 20s sees net decline across all regions; Conference Board: within 3 years, 60-70% of cognitive jobs will enter human-machine collaboration; Maimai: Jan-Jul AI roles +789%, application development roles account for 61%, monthly salary 63,160 yuan.
Take · Three sources point to the same conclusion: not aggregate replacement, but rapid divergence in job geography and age structure—this is both a window for recruitment and skills training, and the real-world foundation for deep-dive Black Swan ② 'cognitive-job replacement triggering regulatory backlash'.
OpenRouter USD-basis reversal: OpenAI spending share overtakes Anthropic for the first time in two years
OpenRouter高管一手数据Confidence HighProgress
For the week of 9/7-13, OpenAI's spending share overtook Anthropic for the first time in two years (Astra 19% > Opus 5's 16%), alongside the metric of China ranking first in Token volume for 20 consecutive weeks.
Take · Cheap models win usage, expensive models win wallets—the two bases must be reported separately; mixing them yields opposite conclusions. When citing this ranking, always specify whether it is the Token-volume basis or the USD-spending basis (routing-only basis).
Market divergence: overseas de-rating, domestic substitution speculation
美股 / A 股证券时报 / 雪球Confidence HighProgress · Day 2
On 9/15, US stocks fell for a second straight day, but the Philadelphia Semiconductor Index rebounded intraday, rising more than 1% (catalyzed by Jensen Huang / Trump publicly opposing AI slowdown on 9/14); China's A-shares saw combined turnover shrink to 1.61 trillion (lowest this year), while STAR 50 was the sole gainer, +1.55%, with a wave of limit-ups in semiconductor equipment / memory / PCB.
Take · Capital exits themes, embraces self-reliance and controllability: under the same fundamentals, overseas assets are priced by discount rate, domestic ones by supply-chain autonomy; the two-market spread itself is a tradable signal.
04RESEARCH2 stories
51 位专家（arXiv 2609.13009）arXiv 9/11（预印本）Confidence HighHard evidence on the false-negative side
51 experts re-evaluate six physics benchmarks: HLE-Physics 47.3% → 78.7% (retained-subset basis), CMT 61.0 → 87.2 (cleanest in the full 50-question audit), PRISM 13.0 → 94.6. Preprint, not peer-reviewed.
Take · Combined with the 'question-selling / false positives' thread since 9/12 into a complete two-way defect map: benchmarks both overestimate (contaminated) and underestimate (grader errors). Raw scores are no longer trustworthy; citations must include harness and data-window basis.
Original ↗ARC-AGI-3 harness gap confirmed: 62.7% vs 99.9%, cost difference 49%
ARC Prize / OpenAI官方 + 新智元Confidence HighHarness confirmed
Astra scores 62.7% ($26,098) under the standard harness and 99.9% ($17,332, saving 49% tokens) under the OpenAI Provider Adapter; Chollet disclosed ARC-AGI-4 will be initiated in Q1 next year (long-horizon continual learning), and ARC-AGI-5 will be initiated around “invention.”
Take · The so-called “benchmark saturation” is actually a harness engineering event: the same model under a different execution framework sees both score and cost change drastically, meaning comparability of evaluation results depends on whether both sides’ harnesses are aligned—execution frameworks must be aligned before procurement acceptance.
05INSIGHT5 stories
Voice Agent industry suite: use a “No.1 + low-cost” combo to make hundred-yuan-level monthly-cost deployment packages.
策略建议①深度版·尚吉策略官Confidence High
Based on the StepAudio 3 (No.1) + U2-Flash (low-cost) combo, build “hundred-yuan-level monthly cost” Voice Agent deployment packages for power / government call centers and in-store sales; delivery projects → industry template library → subscription SaaS. Time window: 4-6 weeks; sign 3 benchmark accounts before Q4.
Take · Leaderboard hype + cost-inflection dividend lasts only 4-6 weeks; the coordinated move is to build an in-house industry evaluation set, hedging the “evaluation trust crisis” as a trust endorsement—sell credibility while selling solutions.
Government-enterprise AI gateway / MCP connectivity-layer business: Zhongyuan Bank's first procurement is the first hard evidence.
策略建议②深度版·尚吉策略官Confidence HighThis week's most worthwhile bet
Using Zhongyuan Bank's first MCP proxy procurement as the anchor, build MCP gateway integration + security assessment for small and midsize banks / city commercial banks; single-point integration → governance platform subscription → voice in industry connectivity standards. Q4 budget season brings centralized procurement; need to build a banking compliance Demo within 1 month.
Take · Procurement shifting down from 'models/applications' to the 'connectivity layer' means the new category has just been independently priced and budgets are not yet fixed; first movers can use a compliance Demo to directly enter the shortlist; Collaboration Framework 3.0 compliance module + 50-320 万 government-enterprise single-pool shared delivery capability.
AEO / multi-model routing service: track validated by capital, no leading player in China yet
策略建议③深度版·尚吉策略官Confidence High
Factory (50 亿) and Profound (18 亿) validate the track; in China, first serve going-global companies with GEO optimization (get products recommended by AI), path is monitoring reports → content transformation → performance-wager subscription, time window 6-12 months.
Take · OpenRouter's USD-denominated reversal confirms the long-term landscape of multi-model coexistence; can combine 'routing + AEO' integrated monitoring; first targeting going-global customers can bypass domestic awareness-education costs for 'AI recommendation optimization'.
This issue's exclusions: three numbers that 'look new but are actually old'
引用纪律深度版·报道台账Confidence HighExcluded items archive
Kimi K3's BrowseComp 91.2% / DesignArena 1388 were recycled from the existing metric basis at the 7/16 release; excluded. Agility Digit 5's 'over $300 million in orders' is a restatement of May SPAC materials; excluded. LMArena's 'three models first to break 1500' is downgraded to 'closed-source leaders first to break 1500' (aggregator-site metric noise). Nex-AGI self-reported scores without independent replication are not reported; OpenAI's rumored $300 million acquisition of Glass Imaging, unconfirmed, is not reported.
Take · At a time of a two-way trust crisis in evaluations, the most valuable move is not to add one more item but to strip out restatements and legacy metric bases—this exclusion log itself is a reusable duplicate-check list.
Three leading indicators: degree of dot-plot hawkishness, whether Anthropic's IPO stays on schedule, and the 9/30 twin inflection points
复核机制深度版·复核机制Confidence High
① Tomorrow morning's FOMC dot-plot hawkishness; ② whether Anthropic's October IPO stays on schedule; ③ the same-day results on 9/30 of the California bill signing and the ARC-AGI-3 leaderboard cutoff. If two of the three reverse, the main logic of 'institutional vacuum + cash-flow pricing' will be systematically revised down at the December review.
Take · Compressing macro, primary market, rules, and evaluations into three falsifiable leading indicators is the most attention-efficient tracker in this report system: no need to reread daily; just recalibrate once every two weeks.