01MODEL RELEASES4 stories
Alibaba releases Qwen3-Max: rises to third on LMArena text leaderboard, surpassing GPT-5-Chat; Hong Kong shares close up 6.4% on the day
阿里（通义千问）安徽画报（引通义千问公众号）2026-09-24；网易·新经济观察 2026-09-24（本机 20 bot《每日AI情报 2026-09-24》转述）Confidence MidMultiple sources agreeAI score 95/100
The company said Qwen3-Max is the largest and most capable model to date; the Qwen3-Max-Instruct preview ranks third on the LMArena text leaderboard, surpassing GPT-5-Chat, while the official version further improves Coding and Agent capabilities. The news drove Alibaba's Hong Kong shares to close up 6.4% at HK$169.6/share, a new high since October 2021, with a total market value of about HK$3.2 trillion.
Take · The only closed-loop sample in this issue: 'model release → same-day secondary-market pricing'—a domestic flagship directly benchmarks against the current OpenAI mainstay on a public human-preference leaderboard for the first time, and capital markets voted with real money that day. External materials and selection tables should use this to recheck the 'Can domestic models enter the main pipeline?' column, but still need to wait for third-party reproduction of the official version.
OpenAI overnight launches GPT-6 Sol / GPT-6 Luna: Luna's per-task cost is about 93%–96% lower than comparison models
OpenAI网易科技 / 21 经济网 2026-09-22～23（本机 20 bot《每日AI情报 2026-09-24》转述）Confidence MidMultiple sources agreeAI score 92/100
OpenAI releases GPT-6 Sol and GPT-6 Luna, both pitched mainly on cost-effectiveness; official estimates put Luna's per-task cost about 93% and 96% lower than comparison models; the report also mentions Texas and that DeepSWE v1.1's max tier scored 66.6% (source basis: vendor estimates and second-hand accounts).
Take · This is the item with the biggest impact on cost accounting this issue: the price cut uses “cost per task” rather than “per million tokens” as its pricing unit, amounting to shifting the buyer’s billing basis from capacity to output. Any budget sheet based on token unit price will be distorted and needs to add “cost per task × first-time-right rate.”
Claude Opus 5.5 pricing goes live: operating costs down about 40% versus the previous generation, input $4 / output $20 per million tokens, cache reads down 60%
Anthropic果壳 2026-09-22（本机 20 bot《每日AI情报 2026-09-24》转述）Confidence MidMultiple sources agreeAI score 88/100
Claude Opus 5.5 pricing details disclosed: operating costs down about 40% versus the previous generation, input $4/million tokens, output $20/million tokens, cache read prices down 60%, with generation speed also further improved.
Take · The 60% cache-read price cut is the largest single cost item in long-context, high-repetition-prefix workloads (exactly our voiceover and daily report pipelines); this more directly rewrites our unit costs than leaderboard rankings, and cache hit-rate assumptions should be reviewed in tandem.
Xiaomi open-sources MiMo-V2.6 (Pro/Flash, native omni-modal) and releases HySparse 2, the core architecture MiMo-V3 will adopt
小米IT之家 2026-09-22/23；36氪「8点1氪」2026-09-22；Latent Space 2026-09-22（本机 20 bot《每日AI情报 2026-09-24》转述）Confidence MidMultiple sources agreeAI score 86/100
Xiaomi open-sources native omni-modal MiMo-V2.6 (Pro/Flash); it calls the release a step in exploring the RSI path. MiMo-V2.6-Pro (1T total parameters / 42B active) scores 46 on the Artificial Analysis Intelligence Index, still trailing Claude Fable 5.1 and GPT-6 Astra. Meanwhile, Luo Fuli announced MiMo-V3 will adopt a new architecture, with the core HySparse 2 cutting prefill FLOPs by 5.02x and shrinking KV cache by 4.5x at a length of 1 million tokens, shifting from block-level to token-level selection and focusing on Agentic reasoning (prefill and KV cache are both on the critical path); Latent Space says its training cost is about $3 million.
Take · Hard efficiency multipliers (5.02×/4.5×) are far scarcer than capability scores—gross margins for long-context Agent are hidden in these two numbers; and if the roughly $3 million training cost can be reproduced, 'training cost' will replace 'leaderboard ranking' as the new competitive benchmark.
02PRODUCT4 stories
HONOR and Alibaba co-build five vertical-domain models: MagicOS 11 launches the industry's first system-level Agent Harness
荣耀 / 阿里36氪 2026-09-22（单源；本机 20 bot《每日AI情报 2026-09-24》转述）Confidence MidSingle-source pending verificationAI score 84/100
Next-generation YOYO: 91.8% overall task accuracy, 90% end-to-end task closure rate, over 100 maximum actionable steps, and average automatic execution time of 3.6 seconds (all vendor-reported; single-source, pending third-party reproduction).
Take · For the first time, a vendor has broken “getting things done” into an assessable four-tuple: closure rate + latency + steps + accuracy—these metrics can be copied directly into our Agent acceptance criteria; but treat them as a target, not a battle report, and they must not go into external materials before independent reproduction.
Microsoft Surface Pro 12 / Laptop 13 with Snapdragon X2 Plus: NPU local inference up to 95% faster, base memory 8GB→16GB
微软IT之家 2026-09-24（本机 20 bot《每日AI情报 2026-09-24》转述）Confidence MidSingle-source pending verificationAI score 78/100
New at Snapdragon Summit: Snapdragon X2 Plus (3nm, Hexagon NPU 80 TOPS), on-device NPU local AI inference up to 95% faster, base memory raised from 8GB to 16GB, first to launch Microsoft Ink Canvas (stylus-driven sketching, annotation, image generation, summarization), available October 13.
Take · On-device crosses the “perceptible performance” threshold: a doubled memory baseline means the resident model tier moves up one level. Our device-cloud split point should be recalculated accordingly—what can be handled locally (initial screening, subtitle correction, sensitive-text preprocessing) should no longer be fully offloaded to the cloud.
T-Head Zhenwu V900 training-inference integrated chip launched: 3x the compute of the previous-generation M890, mass production in 2027 Q1
阿里平头哥IT之家、新浪科技 2026-09-22；网易（引华尔街见闻·硬 AI）2026-09-22（本机 20 bot《每日AI情报 2026-09-24》转述）Confidence MidMultiple sources agreeAI score 76/100
Zhenwu V900 supports 216GB memory, 1200GB/s inter-chip interconnect, native FP8/FP4; compute is 3x that of the previous-generation M890; planned mass production in 2027 Q1; the companion Panjiu supernode will launch in the same period.
Take · The first variable in domestic compute supply is the production ramp of self-developed chips, not data center planning; 2027 Q1 is a verifiable fulfillment anchor; only when the supply chain moves first is 20GW not launch-event rhetoric.
OpenAI officially shuts down Sora API: 2026-09-24 will be the last available day
OpenAIKingy.ai（据 OpenAI 停用公告整理）2026-09-24（另见本机 20 bot《每日AI情报 2026-09-24》多份转述）Confidence HighOfficial announcementAI score 82/100
According to the shutdown notice, Sora API will close on September 24; the consumer website and app were taken offline on 2026-04-26; the official notice reminds users to export assets and developers to migrate integrations before the deadline; unused quota can be transferred to Codex.
Take · Today's hardest news is a 'contract news' item rather than a capability news item: only 5 months separate the consumer shutdown from the API shutdown. Any production line built on a single API (including our video generation line) should use this to conduct a dependency audit and portability design, and write 'lifespan' into the selection table.
03INDUSTRY6 stories
Alibaba Cloud Yunqi compute roadmap: global data centers to exceed 20GW by 2032, training 5T–10T parameter models
阿里云网易（引华尔街见闻·硬 AI）、IT之家 2026-09-22（本机 20 bot《每日AI情报 2026-09-24》转述）Confidence MidMultiple sources agreeAI score 85/100
Group CEO Wu Yongming announced at the Yunqi Conference: Alibaba Cloud's global data center capacity will expand to more than 20GW by 2032, and disclosed the training roadmap for 5T—10T parameter models.
Take · Compute capex is a leading indicator of industry supply; 20GW and 5–10T parameters are order-of-magnitude commitments that will set the floor for unit inference costs over the next 3–5 years. But read it via the dependency graph: the most uncertain link on the critical path is in-house chip mass production (2027 Q1), not power or land.
Three funding items: Snorkel AI $350 million Series E (valuation $3.5 billion), clinical AI assistant Heidi $340 million Series C, Xirang Kaiwu valued at $500 million in two months
Snorkel AI / Heidi / 息壤开物投看·Tocanvc 融资简报 2026-09-23；网易（引蓝鲨财经社）2026-09-23（本机 20 bot《每日AI情报 2026-09-24》转述）Confidence MidMultiple sources agreeAI score 74/100
Data-labeling company Snorkel AI completed a $350 million Series E at a $3.5 billion valuation; clinical AI assistant Heidi raised a $340 million Series C; physical foundation model company Xirang Kaiwu, founded by former Huawei Cloud large-model CTO Li Yin, has raised several hundred million yuan in two months since founding and is valued at $500 million. A total of 11 funding deals were disclosed that day.
Take · Money is flowing to three places—data infrastructure, clinical verticals, and the physical-world foundation—and the cost structure two years from now will grow out of these three. For smaller players, the implication is blunt: don't burn resources competing on general capabilities; put effort into data and scenarios only you have.
CAC adds 3 mobile on-device generative AI service filings: Honor, Xiaomi, StepFun among them
国家网信办每日经济新闻、南方都市报 2026-09-23（本机 20 bot《每日AI情报 2026-09-24》转述）Confidence MidMultiple sources agreeAI score 76/100
Under the Interim Measures for the Administration of Generative AI Services, filing information for 3 on-device services including 'YOYO Claw' was announced; Honor, Xiaomi, and StepFun are among them.
Take · On-device AI enters 'filing as the norm': compliance shifts from 'ex-post ratification' to 'pre-launch', and regulators treat mobile on-device generative AI as a separate category. Device-cloud division plans should put the 'compliance path' before 'performance metrics'—first define tasks that can be compliantly supported, then discuss on-device deployment.
Apple shelves camera-equipped AI wearable pendant project: delayed, no current launch plans
苹果彭博社（古尔曼）经 IT之家转述 2026-09-23（本机 20 bot《每日AI情报 2026-09-24》转述）Confidence MidSingle-source pending verificationAI score 66/100
Bloomberg's Gurman reports Apple's camera-equipped AI wearable pendant project has been delayed, with no current launch plans.
Take · Together with Sora's shutdown, it forms, on the same day, the two sides of 'external dependencies cannot be assumed long-term' (software services can go offline, hardware form factors can be pulled back). Direct product-side conclusion: prioritize reusable software capabilities; hardware narratives do not equal deliverable products.
Regulatory divergence: California signed an AI regulatory executive order on 09-18, while the U.S. federal level sticks to a light-touch approach
加州 / 美国联邦中国青年网 2026-09-23（本机 20 bot《每日AI情报 2026-09-24》转述）Confidence MidSingle-source pending verificationAI score 62/100
California signed an AI regulatory executive order on 2026-09-18; during the same period, the U.S. federal level maintained a light-touch approach, with clear divergence between the two sides.
Take · The state-federal split will first affect site selection, power and water permitting, then flow through to compute supply costs; the practical implication for entities going overseas is to make the stricter side the product premise, not schedule against the most lenient side.
2026 Yunqi Conference closes on 09-24: main theme is AI moving from 'dialogue' to 'doing things' and the Token economy
云栖大会本机 20 bot《每日AI情报 2026-09-24》多家元素转述（安徽画报／IT之家／网易等原始来源）Confidence MidMultiple sources agreeAI score 70/100
Daily intelligence from multiple elements consistently records: this Yunqi Conference's main theme is AI moving from 'dialogue' to 'doing things' and the Token economy; the event also presented Qwen3-Max, 真武 V900 and a 20GW／10T roadmap (see previous item).
Take · 'Token economy' was listed as a key deployment term, echoing the same-day pricing of three models by 'cost per task'—the evaluation system is shifting axes: from benchmark scores to 'how much each closed-loop task costs and how long it takes'.
04RESEARCH3 stories
Thousand-agent orchestration framework Agensh lands on HF Daily Papers: scaling organizational intelligence to 1,024 Agents
Agensh（HF Daily Papers）HF Daily Papers / hf-mirror 2026-09-22（本机 20 bot《每日AI情报 2026-09-24》转述）Confidence MidSingle-source pending verificationAI score 80/100
The paper proposes scaling multi-agent organizations to 1,024 Agents, directly targeting the bottleneck that existing multi-agent frameworks are constrained by the central orchestrator's task allocation and coordination capacity.
Take · This is isomorphic to our daily reality: the real capacity constraint is not how strong any single element is, but the losses in the chain of task assignment, review, and closure. On this basis, 'coordination loss' should be made a first-level metric of our system health (single paper; needs continuous observation).
Research-side dual-system prototype: JEV-as-a-Judge uses a 'decision-type' small model for cheap first-pass screening, then escalates uncertain cases to a stronger model
JEV-as-a-Judge（HF Daily Papers）HF Daily Papers 2026-09-22；Simon Willison 2026-09-21 相关文章（本机 20 bot《每日AI情报 2026-09-24》转述）Confidence MidMultiple sources agreeAI score 78/100
The paper advocates using a 'decision-type' small model for cheap first-pass screening and escalating uncertain samples to a stronger model; combined with Jev-series System One's claim of being 100 times faster and 200 times cheaper than small frontier LLMs, the three independent sources together amount to a layered architecture of 'System 1 cheap and fast (routing/classification/scoring) + System 2 expensive and accurate (deep reasoning)'.
Take · This directly addresses our most expensive steps: judge-panel scoring, image-review gates, and first-pass plagiarism checks. If we still use strong models across the entire chain, that is using System 2 to do System 1's job; wherever one API call can replace one manual ruling, it should be switched out.
RoboFollow paper: Embodied AI success rates systematically overestimated due to 'low scene entropy'
RoboFollow（HF Daily Papers）HF Daily Papers 2026-09-22（本机 20 bot《每日AI情报 2026-09-24》转述）Confidence MidSingle-source pending verificationAI score 72/100
The paper points out that embodied AI success rates are systematically overestimated because evaluation scenarios have 'low scene entropy'—the same type of error as replacing true metrics (lifespan, closed-loop rate) with the most readily available metrics (benchmark scores, demos).
Take · This exposes an old problem in our gating design: available metrics ≠ true metrics. Evaluation sets must be able to falsify themselves (negative controls, scene-entropy checks); otherwise what rises is scores, not capability.
05INSIGHTS6 stories
Simon Willison（Google Gemini TTS）simonwillison.net 2026-09-23（已回读原文）Confidence HighOriginal article, hands-on testedAI score 88/100
Google launched gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts that day, with a built-in 2,000+ voice library, supporting voice customization using just a '30-second audio sample' (one's own voice or a voice one has the right to use); the API can also define multiple characters at once for a full dialogue, each with different voice and style instructions. The author built a BYOK audition console and measured 20 seconds to generate 1 minute 18 seconds of audio at a cost of 2.74 cents (using the more expensive Flash TTS, not Flash-Lite).
Take · The three things our voiceover pipeline needs—custom voice threshold (30 seconds), unit price (about 2.7 cents per minute and a half), generation latency (about 3.7× real-time)—are all delivered at once and can go straight into the cost table for comparison; also a reminder: for any new model, first look at third-party measured tests with unit prices, not just official demos.
Original ↗Simon Willisonsimonwillison.net 2026-09-23（已回读原文）Confidence HighOriginal article, hands-on testedAI score 70/100
This 'tool-type' page uses interactive examples to clearly explain Shadow DOM's style encapsulation, inheritance, slots, parts, and JavaScript access; its generation prompt was just 'Build an artifact to explain shadow roots in CSS with interactive examples' (the page labels it Fable 5.1 Medium).
Take · Shadow DOM is our standard solution for 'embedding third-party components or embedded pages into our own pages without styles clashing'; more valuable is the cost signal—the marginal cost of explainer-type content is already near zero, and the 'explainer pages/teaching artifacts' in our deliverables should be generated by default rather than hand-written.
Original ↗Changing the pricing-unit axis: the first row of the selection table should change from 'benchmark score' to 'unit task cost × task completion rate'.
行业共识（本机 20 bot 多元素同向判断）本机 20 bot《每日AI情报 2026-09-24》多家元素理解与评论（依据 OpenAI／Anthropic／荣耀三方报价口径）Confidence MidMultiple sources agreeAI score 86/100
Multiple elements independently noted: the selling points of the three new models all land on cost per task (Luna lower by about 93%/96%, Opus 5.5 down about 40%), while Honor on the same day quantified 'can get things done' as a 90% closed-loop rate, average 3.6 seconds, and 100+ actionable steps—two seemingly unrelated things converge into the same thing: buyers are starting to settle accounts by 'how much does it cost to complete a task and how many attempts does it take to succeed.'
Take · There is only one action we can directly implement: put 'cost per task × task closed-loop rate' in the first row of any selection or evaluation form, with benchmark scores only as a note, and add a column for cache read unit price; add it to external materials as well, and don't let the other party set the yardstick for us (the conclusion is an element judgment; closed-loop rate and latency are vendor-reported and need third-party reproduction).
'API lifespan' should be written into the selection table: define availability as 'it disappears tomorrow, and I can still open the day after tomorrow.'
行业共识（Sora 关停 + 苹果回撤）本机 20 bot《每日AI情报 2026-09-24》多家元素理解与评论（依据 OpenAI 停用公告、彭博社报道）Confidence MidMultiple sources agreeAI score 84/100
Multiple elements read together the Sora API shutdown (consumer side taken offline in April, API shut down on 9-24, with only a 5-month migration window in between) and Apple shelving its AI pendant: software services can go offline, hardware form factors can retreat, and external dependencies (whether APIs or hardware platforms) do not have long-term certainty.
Take · The most expensive thing is not the API unit price, but the migration cost after weaving it into business processes (a textbook form of the sunk-cost trap). Actionable step: conduct a 'migratability audit' for all external APIs (whether the abstraction layer is independent, whether prompts and workflows can be exported, whether alternative models are in the same tier), and retain a phased right to abandon for production lines.
Redraw the edge-cloud division of labor: first list the 'offloadable checklist', and put compliance before performance.
行业共识（端侧三信号）本机 20 bot《每日AI情报 2026-09-24》多家元素理解与评论（微软新品 / 网信办备案 / 荣耀 Harness 三源）Confidence MidMultiple sources agreeAI score 76/100
3 signals arrived the same day: Microsoft NPU local inference up to 95% faster (memory baseline 8→16GB), China's CAC added 3 mobile on-device generative AI registrations, and Honor's system-level Agent Harness reported a 90% closed-loop rate. For users, the marginal perceived cost of local inference approaches zero—hardware is already paid for, and each run no longer generates a bill.
Take · Directly actionable: list existing steps into an 'offloadable checklist'—subtitle correction, sensitive text preprocessing, and initial screening that can run locally should be offloaded first; heavy inference stays in the cloud. The order is to check the compliance path before performance metrics (on-device registration is already a launch prerequisite). What's saved is not only money, but also data-egress and compliance friction.
AI enters the brand arena: forecast 'public perception' before launch; beyond the cost ledger is a reputation ledger.
观点（内容与品牌）本机 20 bot《每日AI情报 2026-09-24》多家元素理解与评论（Vogue AI 机器人争议）Confidence MidSingle-source pending verificationAI score 64/100
Multi-source records: Vogue runway using an 'undressed AI robot' again sparks public backlash, showing the acceptance red line for AI content in brand spaces is still being repeatedly tested.
Take · For those of us commercializing AI content, the actionable move is to put 'public perception forecasting' upfront in the launch checklist (alongside compliance review); the bigger the company, the more likely it is to skip this in an 'AI equals innovation' atmosphere.
Ecosystem Pulseself-updating
News600Skills72OpenHub44Reddit252026-09-02 → 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 Items4
1. Add 3 columns to the selection table and lock down the definitions: unit task cost × task closed-loop rate, cache read unit price, API lifespan.Change the model-selection template homepage from 'who is strongest' to 'which one can take the dirtiest, largest-volume step off my hands': fixed three hard-field columns—① unit task cost × task closure rate (benchmark scores as notes only) ② cache read unit price (largest single cost item for long-context and high-repetition-prefix workloads) ③ API lifespan and official announcement channels. Current available anchors: Opus 5.5 input 4/output 20 USD per million tokens, cache reads down 60%; Luna per-task cost about 93%～96% lower than comparison models (vendor claim, mark as pending verification). Advanced: run an open-source comparison on own core tasks (MiMo-V2.6-Pro vs current model, fixed task set n≥30, produce capability-cost curve in 7 days).
window Immediate - 2 weeks 2. Do an 'external dependency migratability' inventory: one row each for API/model/platform, noting alternatives and migration hours.Using the Sora API shutdown as a template, list all our single-point dependencies (video generation, voiceover, TTS, image review, gateway, etc.), and annotate each: whether the abstraction layer is independent, whether prompts and workflows can be exported, who the alternative same-tier model is, estimated migration hours, whether assets are self-owned. Criteria are fixed upfront: any 'no alternative and migration effort > 3 person-days' is flagged red, with an alternative path in two weeks. Also set staged abandonment rights for new-direction investments (validation phase→volume ramp phase→scale phase, each stage withdrawable); no one-time heavy bets.
window Immediate - 2 weeks 3. Change the 'human adjudication' step into machine-checkable tiered decisions: cheap first-pass screening by small models, escalate if uncertain.Following the form of JEV-as-a-Judge and System One (claimed to be 100x faster and 200x cheaper than small frontier LLMs), change jury scoring, image review gate, and duplicate-check initial judgment into two tiers: tier 1 uses cheap fast models for routing/classification/scoring and outputs structured results (yes-no gives probability, multi-class gives enum, rating gives score band and criteria), tier 2 handles only uncertain samples. Implementation criteria: tier-1 decisions must retain raw output in ledger and be replayable for recheck; evaluation-type conclusions fix sample size n, and n<30 is always labeled 'for reference only'.
window Immediate - 1 month 4. Produce a 'can-be-pushed-down list': move work that can be done locally back from the cloud, with compliance placed before performance.Map each link of the existing chain by 'compute demand × data sensitivity': subtitle correction, sensitive text preprocessing, initial screening/classification, lightweight orchestration, etc. are prioritized for pushdown to the device/edge (local marginal cost ≈0, avoids data leaving domain), while heavy inference and large-model calls stay in the cloud. Order is fixed as 'compliance path first, then performance metrics'—use CAC on-device filing (Honor/Xiaomi/StepFun already included) to first delineate tasks that can be compliantly hosted; calculate on-device benefits incrementally (bias-variance tradeoff), and do not sacrifice overall architecture reusability for the single selling point of 'localization'.
window Immediate - 1 month