VOL.2026-10-01 · 35 STORIES · SKILLHUB DAILY
SkillHub.news Daily AI
Oct 1, 2026 · #22 · updated 08:00 daily · sources 3
01MODEL RELEASES4 stories
OpenAI releases GPT-6.1 Sol: performance close to GPT-6 Astra, standard token pricing about one-fifth of Astra's.
OpenAI本机20席情报Confidence HighNew topic
20/20 seats report the same: Sol targets coding and computer-use scenarios, with capabilities close to Astra and standard token pricing about one-fifth of Astra's; the ultra-fast version is said to launch 'within days'.
Take · 'Same-generation capability + lower unit price' directly rewrites all Agent per-task cost assumptions, and is the baseline item for all cost-related judgments in this issue.
Google releases Gemini 4 Argon: DeepSWE v1.1 scores 77.9%, priced at $2 per million input / $10 output.
Google格隆汇/财联社/IT之家Confidence HighNew topic
20/20 seats report the same: focused on long-workflow software engineering, legal and financial knowledge work, and cybersecurity defense; DeepSWE v1.1 scores 77.9% (above Claude Opus 5.5's 74.2% and GPT-6 Astra's 74.1%), with a single output cap of about 1 million tokens.
Take · Capabilities and price tag are announced on the same page, so it can go straight into cost tables; output price is 5x input price, with billing centered on 'generation' rather than 'reading'.
Opus 5.5 drives across-the-board cuts of 40—50% in mainstream model prices
Anthropic本机20席情报Confidence MidSingle-source pending verification
20-seat intelligence log (Qianji and Xuanshu seats both report): Claude Opus 5.5 released, driving mainstream model prices down 40—50% across the board.
Take · Frontier models are shifting from 'capability premium' to 'share pricing'; procurement unit prices should not be locked annually and should be renegotiated quarterly.
Sora 2's release date appears questionable in local intelligence
OpenAI（Sora 2）本机20席情报Date questionable · not accepted for now
After verification, the Yanli seat noted: source material claims OpenAI 'officially released Sora 2' today, but most cross-search hits are old 2025 reports, and the Sora 2 page on openai.com also indicates the product is unavailable, so it is handled as 'date questionable · not accepted for now'.
Take · Evidence-discipline trail: momentum must be built on true facts. This item is handled as single-seat pending verification, with no url.
02PRODUCT4 stories
OpenAI launches around-the-clock agent Dots: each instance gets a dedicated cloud VM and can connect to 4000+ apps
OpenAI本机20席情报Confidence HighNew topic
16/20 seats report the same: Dots is based on GPT-6 Astra, with each instance running on a separate cloud VM (Geekbench-leaked specs: AMD EPYC 9V74 / 9 cores / 9.73GB memory / 32GB storage) and can connect to over 4000 apps.
Take · Agents are starting to be metered by 'machine' rather than 'session'; the cost model and security boundary are being rewritten at the same time—always-on means 'one machine stays resident, continuously incurring waiting costs'.
Sora 2 + AgentKit + Codex official version launched in a same-day bundle (including Sora 2 API)
OpenAI本机20席情报Confidence MidRelease date in doubt
20/20 seats report the same bundled release: Sora 2 is officially called the 'ChatGPT moment for video generation'; AgentKit toolchain, Codex official version, and Sora 2 API were released at the same time; however, Sora 2's exact release date is disputed (see disputed model-category entry).
Take · The unit of competition has upgraded from 'a single model' to 'a reusable tool asset bundle,' which is most likely to solidify into a standard component.
Meta personal assistant Muse: tops U.S. App Store free chart less than two weeks after launch; estimated downloads exceed 3.4 million
Meta36氪/Sensor Tower/华尔街见闻Confidence MidUpdate · Day 2 (09-30 already reported its previous round of downloads)
20/20 seats jointly report: Sensor Tower estimates cumulative downloads exceeded 3.4 million as of September 24; Tencent, ByteDance and Alibaba all enter, with Manus and OpenAI following.
Take · Download volume is a customer acquisition metric, not value delivery; meanwhile TechCrunch points to ugly consumer AI economics, forming a 'hype vs ledger' comparison sample.
DoorDash官方·TechCrunch AIConfidence High
TechCrunch reports: DoorDash launches an AI agent that can complete food ordering via text-message conversation (headline taken from the original ledger).
Take · Another piece of evidence that 'transaction execution rights' are shifting from App pages to conversational interfaces, on the same track as the platform access fight.
Original ↗03INDUSTRY7 stories
AMD to acquire Fei-Fei Li-founded World Labs for about $8.2 billion in an all-stock deal.
AMD观察者网/21世纪经济报道/联合早报/新浪财经Confidence HighNew topic
20/20 outlets report the same: all-stock deal, expected to close by end of 2026, pending regulatory approval; AMD's second-largest acquisition ever; Fei-Fei Li appointed chief scientist.
Take · Chipmakers shift from 'selling compute' to 'buying capability'; world models/spatial intelligence move from paper narrative to capacity narrative, and the independent valuation anchors of second-tier model teams begin to loosen.
Anthropic prospectus revealed: FY2025 revenue $4.59 billion (+1088%), operating loss over $8 billion
Anthropic财联社/新浪财经Confidence Mid
20/20 outlets report the same: nearly a quarter of revenue comes from two customers, most major customers have not signed long-term contracts; compute/cloud/infrastructure obligations total $518 billion; separately signed a 7-year $11.6 billion Akamai procurement (more than 6x the $1.8 billion deal in May).
Take · Xuanzhu Yixi correction: $518 billion is a cumulative obligation over many future years, not a single-year budget (the source draft's 'next year' framing is incorrect). Revenue is revocable, spending is irrevocable—a classic maturity mismatch.
OpenAI plans to raise at least $30 billion in new funding, targeting a $1.4 trillion valuation
OpenAI界面新闻Confidence Mid
19/20 seats concur: possibly bridge financing before IPO; a March round this year already secured $122 billion in committed capital at an $852 billion valuation.
Take · A valuation move from $852 billion to a $1.4 trillion target in six months is the main anchor for AI asset prices this round, directly determining upstream and downstream costs and talent flows.
Micron FY2026 record: revenue $133.188 billion (+256.33%), net profit $84.969 billion (+895.07%)
美光IT之家/新浪财经Confidence High
18/20 seats concur: Q4 single-quarter revenue topped $54.2 billion; it has mass-produced the industry's first 512GB high-density DDR5 RDIMM and forecasts continued memory supply tightness in 2027–2028.
Take · AI demand's bottleneck is shifting from compute to storage; the default assumption that 'token will inevitably get cheaper' must be recalculated using two factors: 'unit price down 50% / call volume up 3x'.
Two-track regulation: FTC opens industry-wide probe into Anthropic, OpenAI, and others; Connecticut AI law takes effect October 1.
监管（FTC / 康涅狄格州）IT之家/Superpower DailyConfidence High
18/20 outlets report the same: senior FTC official says investigation covers the entire industry; Connecticut legislation adds consumer and worker protection provisions, effective October 1.
Take · Compliance is shifting from 'post-hoc fines' to 'ex-ante rules,' from a cost item to an access requirement.
ElevenLabs valuation rises to $22 billion, doubling in less than eight months.
ElevenLabs财联社Confidence MidSingle-source pending verification
6/20 outlets report the same: scaled deployment in enterprise customer service and other scenarios is the main driver, with capital accelerating into the AI voice agent track.
Take · Voice is one of the few scenarios with proven scaled paid adoption, and can be tracked separately as a monetization bellwether.
Reddit官方·TechCrunch AIConfidence High
TechCrunch reports: Reddit halts RSS subscriptions and ends public API access, citing AI bot scraping (headline taken from original ledger entry).
Take · Content sources continue to escalate their blocking of AI, directly raising the cost of collecting and training on public-domain corpora.
Original ↗04RESEARCH4 stories
Org-Agent: From personal assistant to organization-level agent.
Hugging Face Daily PapersHF 镜像Confidence MidSubmission date/listing batch noted.
The paper proposes that organization-level agents require cross-user interaction/decision-making and cross-user memory/knowledge accumulation (listed in the 2026-09-29 batch).
Take · It points to the same proposition as the same-day always-on agent product route (inheritable cross-session memory), but lacks scaled validation.
Can Agents Design Libraries for Agents?（LibraryDesignBench）
Hugging Face Daily PapersHF 镜像Confidence MidSubmission date/listing batch noted.
Research indicates agents tend to 'reimplement rather than reuse,' causing codebases to keep expanding and dragging down subsequent agents (listed batch 2026-09-29).
Take · If no one archives and deduplicates Agent artifacts, the knowledge base will turn from an asset into a liability; production environments should roll out artifact deduplication and archiving gates as soon as possible.
Routing Should Pay for Itself: Router Cost Constraints and Sparse Supervision
Hugging Face Daily PapersHF 镜像Confidence MidSubmission date/listing batch noted.
Discusses that model routers must earn back their overhead, involving routing efficiency under sparse supervision (listed batch 2026-09-29).
Take · The routing layer is beginning to be required to 'pay for itself,' in line with inference cost becoming the only sensitive variable.
Apple ML官方·Apple MLConfidence Mid
Apple machine learning research: a systematic study of the effectiveness–fluency trade-off in LLM conditioning (title taken from the original ledger).
Take · The trade-off between capability and fluency is still under systematic study, suggesting the product-side prior of 'stronger is better' should be discounted.
Original ↗05INSIGHTS6 stories
Procurement price comparisons must switch metrics: from 'unit price per million tokens' to 'unit task cost (¥/条)'
观点（本机20席情报·巽风／乾机）20 席情报Confidence Mid
Unit price falls 80% while agent steps rise 5x, but the total bill rises; Decisions API shifts high-frequency, low-judgment nodes from per-token to per-call billing—when granularity changes, gross margin structure changes.
Take · The first cost-cutting action is not switching to a cheaper model, but cutting unnecessary decision nodes.
The main competitive axis shifts from 'model scores' to 'cost per task × task completion rate'.
观点（本机20席情报·玄枢／渊策）20 席情报Confidence Mid
After the three players' capabilities are pulled to the same level, the only difference left is 'who gets called by default'; key metrics should shift to D7 retention, gross profit per user, and task completion rate, not leaderboard scores.
Take · Any project whose core assumption is 'we have a better model' should change that assumption today.
3.4 million downloads is demand awakening, not demand validation.
观点（本机20席情报·惊蛰／渊策／坎流）20 席情报Confidence Mid
First validate three things at low cost: D1/D7 retention, per-user gross margin, and task completion rate; before retention data is in, buying downloads with subsidies equals doing user education for the entire industry.
Take · Inference costs rise linearly with usage time while subscription revenue stays fixed; the more successful a user's usage, the more you lose.
Compliance is shifting from a 'cost item' to an 'entry requirement': make compliance delivery packages standard components.
观点（本机20席情报·厚坤／悦泽）20 席情报Confidence Mid
Auditability, explainability, and accountability are shifting from bonus items to entry requirements; recommend making 'data flow diagram + log retention + human review points' standard components.
Take · Teams that pass review on the first attempt have lower long-term customer acquisition costs and more stable long-term contract credit.
Entry points will change, assets won't: turn 'task chains' into reusable assets
观点（本机20席情报·明见／玄枢／润知）20 席情报Confidence Mid
Whichever assistant wins, the asset set of 'memory structure + permission rules + failure rollback' can be ported; change 'full delegation' to 'tiered authorization + one-click rollback'.
Take · The breakthrough for delegation is not capability, but failure being visible / retractable / takeover-ready.
Compute 'long-term contract': asset or debt? First run a 'supply cutoff / 2x price increase' stress test.
观点（本机20席情报·厚坤／乾机／坎流）20 席情报Confidence Mid
It is recommended to stress-test all business lines that rely on a single model supplier; any business dependent on subsidized pricing from leading model companies should reserve a fallback that 'still holds if prices rise 3x'.
Take · The maturity mismatch between capacity commitments (multi-year, rigid) and customer contracts (one-year, cancellable) is the most expensive bet of this round.
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 Items3
1. Shift Agent procurement and pricing anchors from "token unit price" to "cost per task."Within this week, re-account all in-use model calls by "actual cost per completed task" (including retries and tool calls), then use that to renegotiate or reroute suppliers; external pricing should also switch from hourly/per-use to per-task.
window 1-2 months 2. Build a "trustworthy" trio for always-on Agents.On your in-house agent line, add the three actions—failure visibility, revocability, and takeover—and require double confirmation for irreversible operations; validate with D7 retention and task completion rate, not call volume.
window 3-6 months 3. Two-factor sensitivity table for compute and storage budgetsCalculate one budget version for each of two scenarios: 'unit price down 50% / call volume up 3x' and 'tight storage supply, rising quotes'; for projects reliant on large storage/bandwidth, assess stockpiling pace and long-term contract terms, preferring to pay a small redundant holding cost.
window This quarter