VOL.2026-10-20 · 29 STORIES · SKILLHUB DAILY
SkillHub.news Daily AI
Oct 20, 2026 · #38 · updated 08:00 daily · sources 4
01MODEL RELEASES3 stories
V4.1-Flash-1017 average cost per task 3.3→1.3 yuan: unit price unchanged, shorter trajectories
DeepSeek机器之心 + 财新Confidence HighProgress · Day 3
763B parameters; Terminal-Bench 4.0 31.2→42.3 (+11.1 points), ProgramBench 20.3→29.1, DeepSWE v1.1 74.2%→77.0%; API unit price unchanged, while max-tier average cost per task fell from about 3.3 yuan to about 1.3 yuan (-61%), with self-developed Search + Fetch retrieval API also opened
Take · Unit price unchanged while per-task cost drops 61%—there is only one explanation: fewer tokens consumed to complete the same task (shorter trial-and-error trajectories, fewer environment round trips). The cost reduction occurs at the trajectory and retry layer, not in list price; the three scores are vendor-reported on its own benchmark, and cross-vendor comparisons require fixing the task set and scoring script
MiniMax M3.1 Flash: inference nearly 3× faster, cost down 70%
MiniMax财联社Confidence MidSingle-source pending verification
Official claims: inference nearly 3× faster and cost down 70%, with generational gains in Agentic tasks / coding / Computer Use / cybersecurity; previously, under anonymous identity Space Bunny, it had already run over 200T tokens of real traffic; Pro version preview is imminent
Take · Scale anonymously first, then launch under its real name—real traffic during the anonymous period is its free evaluation; if the reduction lands, second-tier model vendors will see further gross-margin pressure, opening a bargaining window for buyers
NVIDIA dual-tier open-source weights: Nemotron 3.5 Super VL (120 billion) and edge-side Bolt
英伟达IT时代网 + TechCrunchConfidence HighNewcomer
Nemotron 3.5 Super VL is a 120-billion-parameter model for always-on enterprise agents; the company says its instruction-following and tool-calling accuracy approach those of the larger Nemotron 3 Ultra; Bolt can run on RTX Spark AI PC and DGX Spark.
Take · Open-sourcing both enterprise and device-side tiers on the same day effectively embeds the two touchpoints—'always-on cloud' and 'offline on-device'—into its own toolchain; the moat shifts from chips to deployment validation and the default stack.
02PRODUCT RELEASES3 stories
AXP fills in engineering metrics: million-level registrations and 100,000-level concurrent online, but the hard constraint is a '2.5-hour disconnection'.
紫金山实验室 / 中国移动码农财经·今日头条Confidence MidProgress · Day 3
Based on a deterministic network: million-level agent registrations, 100,000-level concurrent online, cross-domain addressing over 4000 kilometers, and end-to-end task time reduced by more than 20%; Academician Liu Yunjie says that under the traditional internet architecture, agents working continuously for about two and a half hours will disconnect from the network.
Take · The ceiling for agents is not intelligence but session lifespan—long-task agent failure rates have been systematically misattributed (single source, pending verification, but this is the most engineerable item of the day).
ATI trusted infrastructure for agents: reuses DNS/PKI to issue 'ID cards' to agents.
CNNIC × 阿里云素材补充件①Confidence MidSingle-source pending verification
Appeared the same week as AXP, reusing existing DNS and PKI systems to issue cross-organization mutually recognized identity credentials for agents.
Take · The road (AXP) and license (ATI) appeared the same week—agents' “address layer + identity layer” is starting to become infrastructure, and the closed-source protocol camp must answer interoperability questions.
Databricks官方博客 / 官方视频Confidence High
The official videos “Introducing Agent Memory on Databricks” and “Building Production-ready AI Agents” launched simultaneously, with Agent Memory released as a production-grade agent memory component; the company also announced doubling investment in the New York AI talent center.
Take · Memory has moved from a “prompt trick” to a “platform component”—differentiation in agent engineering is shifting from prompt to state management and retrieval.
Original ↗03INDUSTRY6 stories
美银 / 微软 / Nscale新浪财经 + Reuters/FT/BBCConfidence High
BofA estimates hyperscalers' FCF will fall below -$100 billion over the next 12 months, with capex consuming 95% of operating cash flow; over the same period, NVIDIA/Micron/Broadcom combined FCF is near +$500 billion; in the same week, Microsoft notified termination of its Essex (Loughton) data center anchor lease, and Nscale's up to $35 billion listing plan suffered a setback.
Take · The -$100 billion is an aggregate figure, not a universal one (Microsoft is the only one of the five largest players with positive FCF in the current period), so accept it on an aggregate basis but reject it for individual-stock conclusions; the real leading indicator is lease termination—the first demand-side evidence of 'withdrawing commitments'.
Original ↗Tenstorrent's $1.5 billion Series E at a $6.2 billion post-money valuation
TenstorrentIT之家 + TechCrunchConfidence High
Led by Trace Capital and Innovation X, doubling the roughly $2.7 billion valuation from its late-2024 Series D; proceeds will fund large-scale mass production and global commercial expansion, with Jim Keller emphasizing commitment to an open, scalable architecture.
Take · In a selloff week for the compute sector, the 'second supplier' can still command double the pricing—the premium for selling hardware alone is being continuously compressed; a portable open architecture is the real pricing anchor.
英伟达 GTC 柏林界面新闻 + NVIDIA BlogConfidence High
10/20—22, 100+ sessions; Jensen Huang speaks on the 21st, focusing on Rubin platform deployment validation and about $2 billion in European sovereign AI investment; concurrent blog posts present two implementation cases: German industrial AI partnership and media agentic security.
Take · The narrative is shifting from selling compute to selling capacity—if the pricing unit shifts from per-token to per-MW accordingly, buyers' budget structures will need to be reworked.
Original ↗亚马逊TechCrunchConfidence MidHeadline-level evidence
TechCrunch reported on 10-19 that Amazon launched the TorchNeuron developer tool, described as its biggest offensive against NVIDIA to date.
Take · The competitive focus shifts from hardware to developer migration cost—real replacement only comes when existing PyTorch code can run with a few lines changed.
Original ↗Salesforce: customer service traffic from personal Agents will exceed the amount they offset; simultaneously pushes Slack Frontline
Salesforce官方新闻稿Confidence High
Official says customer-service traffic created by personal agents will exceed the amount they deflect for users; the same day, a Slack version for frontline shifts/sites/teams (Slack Frontline) was launched.
Take · Agents do not eliminate customer-service demand; they shift the entry point from 'people seeking customer service' to 'agents seeking customer service'—support reps' work becomes handling agent requests, and both SLA and ticket structure must change.
FSB final guidance document on AI in finance: 12 operational principles
FSB官方（fsb.org 原文交叉）Confidence High
Published 10-19, covering organizational governance, risk assessment, data management, human oversight, cybersecurity, and third-party risk management; not legally binding, but a bellwether.
Take · The 12 items can be condensed directly into a one-page bid self-assessment form—non-binding, but likely to become a buyer due-diligence checklist; early self-certification is a signaling advantage.
04RESEARCH2 stories
All three new scores are vendor self-reported from its own question bank: Terminal-Bench 4.0 42.3 / ProgramBench 29.1 / DeepSWE 77.0%
评测口径深度版复算 + 情报交叉Confidence MidDefinitional caveat
Terminal-Bench 4.0 31.2→42.3, ProgramBench 20.3→29.1, DeepSWE v1.1 74.2%→77.0%; all three readings come from the same vendor, same batch of releases, with samples from its own question bank.
Take · Before cross-vendor comparisons, the task set and scoring script must be fixed; otherwise it is only correlation, not causation. Falsifiable criterion: if “per task” truly becomes a pricing unit, mainstream API pricing pages should add “per task/per workflow” tiers within 30 days.
OpenAI官方博客Confidence High
10-20 official release: coding capabilities for low-resource programming languages (openai.com/index/low-resource-programming-languages); same day, Databricks launched Agent Memory, a production-grade memory component.
Take · Coding capabilities spread to long-tail languages: industrial scripts, legacy systems, and small-language ecosystems are the next deliverable scenarios—and the hardest corners for domestic substitution to cover.
Original ↗05INSIGHT5 stories
Pricing unit shift = contract structure shift
前瞻深度版研判Confidence MidAssessment
DeepSeek per-task -61% and MiniMax cost -70% emerged on the same day—buyers are starting to account by “how much it costs to finish one job,” not by token unit price.
Take · When buyers budget on a 'per task' basis, suppliers must compete on equal task costs; the industry slides toward Bertrand-style price competition, and profit shifts to those who can get tasks done. For buyers, this is a 3—6 month window; afterward, the benefit is clawed back in one go at the next price adjustment.
'Session lifespan' is a new billable category
悦泽 / 润知 / 渊策 / 艮山四路情报同日独立提出Confidence Mid
A 2.5-hour network outage means the failure rate of long-task agents has been systematically misattributed. Deliverables: session survival curves (stratified by task type, with three checkpoints at 30 minutes/2 hours/8 hours), breakpoint location distribution, automatic checkpointing and resume after disconnection, heartbeat liveness checks, and visualization of '70% complete' after recovery.
Take · For long-process customers in patrol inspection, reconciliation, and compliance review, this is a must-have; acceptance metrics should be 'task survival rate × average recovery cost,' not 'model score.'
The 3320 亿 figure does not match the stated basis: the implied four-year CAGR is about 99%, a 580 亿 gap from the labeled 107.6%.
坎流情报 / 36氪恒等式复算Confidence HighRecalculation
36氪研究院 gives 2025: 212 亿 → 2029: 3320 亿, CAGR 107.6%; but 212 → 3320 is 15.66 times, and the implied four-year CAGR is about 99%. If it really grew at 107.6%, 2029 should be about 3935 亿, a gap of 580 亿 (17%).
Take · Based on identity recalculation, the labeled 107.6% is judged as mixed definitions; TAM serves only as a directional anchor, not a budgeting basis.
Open-source weights are a distribution channel, not charity.
生态深度版研判Confidence High
NVIDIA released enterprise-grade Super VL and edge-side Bolt on the same day; Amazon used TorchNeuron the same day to capture developer migration costs.
Take · Whoever controls 'runs after changing a few lines of code' controls the next layer's default stack—this layer is more expensive than the weights themselves.
Optimizing networks for agents may make things worse for others.
MIT Technology Review外媒Confidence MidHeadline-level evidence
MIT Tech Review 10-19 article: Optimizing networks for agents (dedicated channels, deterministic latency, identity systems) may come at the expense of regular users' experience.
Take · This is the same coin as AXP/ATI—whom infrastructure is optimized for is a public choice that must be priced in at the standards stage.
Ecosystem Pulseself-updating
News600Skills67OpenHub44Reddit302026-09-02 → 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. Switch the quoting/procurement metric to “full cost per task”Switch the quoting/procurement metric to “full cost per task”; change both external quoting and internal procurement KPIs from “price per million tokens” to “full cost per task (including retries, manual fallback, and acceptance rework)”; lock in three contract elements: a per-task cost cap + advance notice period for price increases + same-tier benchmarking (MFN) clause. Path: first run a 7-day comparison using 20—30 real tasks of your own (20 runs each for the new and old models, recording only first-pass rate and manual rework minutes); if it does not beat the current production baseline, do not switch → then quote on a tiered basis by results (volume/closed-loop rate/share of labor savings).
window Effective immediately, 3—6 months; criterion: if mainstream API pricing pages still show no “per-task” tier within 30 days, the benefit is only a billing construct—scale back investment 2. Agent “session lifetime” engineering packageDeliver an “always-connected agent” for long-workflow customers: automatic checkpoint on disconnect, resume from checkpoint, heartbeat liveness checks, human takeover points, and visualization of completed progress; acceptance metric is “task survival rate × average recovery cost,” not model scores. Approach: first run survival-rate tests on 20 long tasks of 2 hours or more, plot the breakpoint distribution (most likely at 2—2.5 hours) → treat checkpoint density as an optimization variable to find the intersection of “reload cost vs checkpoint overhead” → turn it into a standard deliverable, benchmarked against SLA commitments.
window Effective immediately, 2 quarters; within this window, third-party access terms for AXP/ATI have not been disclosed—build in-house first and do not bet on a single platform 3. FSB 12 items → financial-client bid appendixCompress the FSB 12 principles into a one-page “12 questions before launching an AI application” self-assessment form, as a standard appendix for financial/medical client bids and due diligence. Approach: six-column mapping (governance/risk assessment/data/human oversight/cybersecurity/third party) → for each item, provide auditable evidence (who signs, what trail is kept, how often it is reviewed) → upgrade into a quarterly review report, forming a governance evidence chain and turning compliance from a cost item into an entry barrier.
window 1—2 quarters; translating soft law into national rules usually takes 3—6 months; waiting for mandatory implementation before catching up will double the cost