Boss lens: more tokens ≠ higher bills—bake price wars and thrift into AI budgets, don’t underwrite pure hypergrowth (inference).
AI NEWS
Today’s AI News
Chinese and English are the same day. This page lists every item, each with the original link, the author, and the engagement.
Latest day
November 7, 2026
11-06 07:50 – 11-07 07:50 (UTC+8)
Top 3
Doubao Work personal plan drops 5-hour short-cycle quota from Nov 6; weekly quotas stay
Boss lens: short-cycle gates are gone—manage weekly caps so peak work isn’t clipped mid-shift (inference).
Boss lens: ad/creative video unit cost and revision speed just moved—budget draft→upscale workflows before scaling spend (inference).
All
Ramp: enterprise AI token spend up only ~1.5% across Aug–Sep (vs ~+42% MoM in July) while usage rises; Anthropic IPO talk ~$1.2–1.3T
Boss lens: more tokens ≠ higher bills—bake price wars and thrift into AI budgets, don’t underwrite pure hypergrowth (inference).
CollapseRamp: enterprise AI token spend up only ~1.5% across Aug–Sep (vs ~+42% MoM in July) while usage rises; Anthropic IPO talk ~$1.2–1.3T
Per TechCrunch citing Ramp (via @dotey): firms still burn more tokens, but spend nearly flatlined—July MoM ~+42%, August ~+1%, September ~+0.2%; roughly early Aug–mid Oct usage ~+65% vs spend ~+1.5%. Drivers include price cuts, buyer thrift/routers, and cheaper alternatives. Coverage also has Anthropic IPO chatter toward ~$1.2–1.3T from ~$2T talk, and OpenAI’s round near ~$1.25T with a smaller raise target. Ramp’s sample skews tech—not a full-market census.
Doubao Work personal plan drops 5-hour short-cycle quota from Nov 6; weekly quotas stay
Boss lens: short-cycle gates are gone—manage weekly caps so peak work isn’t clipped mid-shift (inference).
CollapseDoubao Work personal plan drops 5-hour short-cycle quota from Nov 6; weekly quotas stay
Per IT Home citing Era Finance: from Nov 6, Doubao Work’s personal plan removes the 5-hour short-cycle quota. Users can pace usage inside the weekly allowance; weekly caps and other plan rules still apply. Check remaining quota under Settings → Subscription & quota. Doubao Work targets productivity workflows—research, writing, data jobs, and light app building via natural language. Other CN outlets match the same framing.
Google Gemini Omni 1.2 Flash: up to 15s video, ~30% cheaper at $0.07/s (720p), multi-turn edits; GA on API/Cloud
Boss lens: ad/creative video unit cost and revision speed just moved—budget draft→upscale workflows before scaling spend (inference).
CollapseGoogle Gemini Omni 1.2 Flash: up to 15s video, ~30% cheaper at $0.07/s (720p), multi-turn edits; GA on API/Cloud
Per Google’s blog (via @dotey): Gemini Omni 1.2 Flash makes native clips up to ~15s (vs 10s on 1.1), about 30% cheaper at ~$0.07/s for 720p, with multi-turn edits and stronger reference adherence. It is GA on the Gemini API and Google Cloud, with Vids/Flow/Ads paths noted. Roughly ~$1 for a 15s 720p clip at list pricing. Video-gen product update—not a new frontier LLM.
Meta Muse launches Spaces: describe a custom interactive space inside the app; pin to sidebar or tabs
Boss lens: pocket DIY mini-apps got easier—benchmark your internal light tools/knowledge hubs against ‘describe it, get it’ (inference).
CollapseMeta Muse launches Spaces: describe a custom interactive space inside the app; pin to sidebar or tabs
Per Meta’s Nov 6 blog: Muse adds Spaces—describe an idea (gift tracker, planner, small-business order board) and Muse builds an interactive space you can refine, pin to the sidebar/main tabs, hide, or schedule. Connected apps can keep spaces updated in the background. Spaces roll out wherever Muse is available; specialized models trained with Muse Wonder learnings are also rolling into Muse/Meta AI. Consumer personal-agent UX, not an enterprise suite.
OpenAI BioBasicsBench: 100 wet-lab routine tasks; GPT-6 Astra ~66.7% vs GPT-5.6 Sol ~43.8% at xhigh
Boss lens for bio/lab ops: agent reading of raw experiment evidence is now scored—require control literacy before production use (inference).
CollapseOpenAI BioBasicsBench: 100 wet-lab routine tasks; GPT-6 Astra ~66.7% vs GPT-5.6 Sol ~43.8% at xhigh
Per OpenAI (Nov 6): with biotech partner Ligo, BioBasicsBench scores agents on 100 routine wet-lab analyses—raw gels, plates, microscopy, exports—testing interpretation, error spotting, and knowing when evidence is insufficient. At xhigh reasoning, GPT-6 Astra averages ~66.7% vs ~43.8% for GPT-5.6 Sol. OpenAI also shows remaining failure modes (control misreads, colony overcounts). Benchmark release, not a new model launch.
Mondo Robotics’ Beni tops ~$4.6M on Kickstarter; new round targets ~$1.2B valuation (LeiPhone)
Boss lens: embodiment can start as a sellable niche (follow-cam), not full humanoids—Kickstarter is a demand signal (inference).
CollapseMondo Robotics’ Beni tops ~$4.6M on Kickstarter; new round targets ~$1.2B valuation (LeiPhone)
Per LeiPhone (Nov 6): Mondo Robotics’ Beni camera robot finished Kickstarter above ~$4.6M with 7,300+ backers (passed ~$1.25M in ~10 hours). A new financing round is said to target ~$1.2B valuation after earlier ~$7B talk filled quickly. Beni is a wheeled all-terrain follow-cam robot aimed at outdoor/daily capture, with consumer pricing around ~$799. Figures are press reporting on fundraising progress, not audited closes.
Claude Code cloud sessions: Pro/Max/Ultra can auto-detect project deps and required API keys
Boss lens: cloud agents that self-wire deps speed demos—tighten secrets policy before wide rollout (inference).
CollapseClaude Code cloud sessions: Pro/Max/Ultra can auto-detect project deps and required API keys
Per TestingCatalog (@testingcatalog) on Nov 6: on Pro/Max/Ultra, Claude Code cloud sessions can figure out project dependencies and required API keys so a repo can run in the cloud. Demo video framed as “Claude in the cloud.” Product observation post—confirm exact rollout notes in Anthropic docs; no price-change claim here.
Airbnb plans an AI travel agent in H1 2027 (earnings remarks via TechCrunch et al.)
Boss lens: booking platforms are merging advise+book into one agent path—map your sales/support funnel to the same shape (inference).
CollapseAirbnb plans an AI travel agent in H1 2027 (earnings remarks via TechCrunch et al.)
Per TechCrunch/The Verge coverage of Airbnb’s Q3 2026 remarks: the company targets an AI travel agent in H1 2027. Framing emphasizes a generative travel UI with maps, photos, comparisons, calendars, and itineraries—not a plain chatbot—to help guests explore, decide, and book. Full feature/pricing details are not yet fully disclosed. Forward-looking guidance until a product launch.
US DOJ launches Super Intelligence Task Force (DOJ-SITF) to coordinate SI enforcement
Boss lens for US exposure: SI misuse now has a dedicated DOJ task force—tighten use-case and audit narratives (inference).
CollapseUS DOJ launches Super Intelligence Task Force (DOJ-SITF) to coordinate SI enforcement
Per DOJ OPA (Nov 5; recirculated in-window): the Department launched a Super Intelligence Task Force (DOJ-SITF), led by Associate AG Stanley Woodward and Associate DAG Stephen Weiss, to counter misuse of “super intelligence” by foreign actors, TCOs, and fraudsters, and to coordinate with the White House SI Force. Organizational announcement—not a full rulebook.
USA Today Co.: AI search hits traffic; Q3 digital ad revenue ~$75.0M (~-13.9% YoY)
Boss lens: search-ad publishers are being skimmed by AI answers—optimize to be cited/summarized, not only clicked (inference).
CollapseUSA Today Co.: AI search hits traffic; Q3 digital ad revenue ~$75.0M (~-13.9% YoY)
Per Q3 2026 earnings-call coverage of USA Today Co. (TDAY): management tied softer search traffic/pageviews partly to AI-driven search shifts. Digital advertising revenue was about $75.0M, ~13.9% YoY lower; indirect ads (programmatic/sponsored recs) fell ~32% YoY. The company also discussed AI efficiency and ~$100M annualized cost actions. Earnings-call figures—confirm details in filings.
Perplexity launches Teammates: always-on agents for ongoing work (Max)
Boss lens: chat becomes always-on watchers—pilot 1–2 measurable standing workflows before seat sprawl (inference).
CollapsePerplexity launches Teammates: always-on agents for ongoing work (Max)
Per Perplexity’s product pages/blog and The Verge: Teammates are always-on work agents that plan, monitor changes, and act proactively—optionally coordinating other agents on ongoing jobs. Initially available to individual Max subscribers on web; Max list pricing is about $200/month or $2,000/year. Product launch—confirm enterprise seats/permissions in docs.
Mistral Studio public preview: Managed Agents and Apps for hosted agents and deploys
Boss lens: another EU-leaning agent+app host option—audit tenancy, logs, and exit paths in the preview (inference).
CollapseMistral Studio public preview: Managed Agents and Apps for hosted agents and deploys
Per Mistral docs and release notes (~Nov 5, recirculated in-window): Studio puts Managed Agents and Apps into public preview. Define agents in Python, run in-process, self-hosted, or Mistral-managed with durable background execution and Studio tracing; scaffold/deploy Apps via CLI to Mistral Cloud. Preview platform features—check console for quotas/pricing.
Cursor cloud agents: describe an app, provision DB/auth, and deploy a live URL
Boss lens: internal tools get closer to ‘describe → live’—pair speed with data-class and release gates (inference).
CollapseCursor cloud agents: describe an app, provision DB/auth, and deploy a live URL
Per Cursor’s Build Live Apps changelog/docs: paid users with cloud-agent access can ask an agent to build a web app from natural language, provision Postgres and auth when needed, and publish a live URL after security checks. Framed as getting past localhost to a hosted app. Confirm plan gates and security checklist in Cursor docs.
STO Express partners with Qwen Office to push Agent rollout to end-network outlets
Boss lens: logistics majors push agents to outlets—scale lives at the edge, not HQ demos; map your store/outlet SOPs (inference).
CollapseSTO Express partners with Qwen Office to push Agent rollout to end-network outlets
Per LeiPhone / China Transport News / Sina Finance (Nov 6): STO Express and Qwen Office announced a strategic partnership to build AI office workflows for core express scenarios and extend Agent capabilities from HQ into end-network outlets nationwide. Corporate MoU-style coverage—rollout timing, spend, and pilot scope are not fully disclosed.
Community claim on DeepSeek: ~1.54M DAU, ~92T tokens/day, ~35k third-party plugins (unverified)Unconfirmed
Boss lens: high plugin penetration would be a strong ecosystem signal—but treat figures as unverified until official (inference).
CollapseCommunity claim on DeepSeek: ~1.54M DAU, ~92T tokens/day, ~35k third-party plugins (unverified)
Per @mylifcc (Nov 6) relaying community figures: DeepSeek ~1.54M DAU, ~92 trillion tokens/day, ~35k third-party plugins, with ~60% of active users said to have built/installed plugins. No official report link in-post; no matching official press note found in quick checks. Marked unverified community hearsay—signal only, not audited metrics.
Zhipu ZCode weekend airdrop: ~500M tokens of GLM-5.3-Flash for a limited window
Boss lens: use weekend free tokens for batch evals—don’t staff capacity on expiring promo quota (inference).
CollapseZhipu ZCode weekend airdrop: ~500M tokens of GLM-5.3-Flash for a limited window
Per @NFT_Chen (Nov 6): Zhipu’s ZCode client is airdropping about 500 million tokens of GLM-5.3-Flash for a weekend window (said to activate 23:00 that night through Monday 09:00). Promo/quota post—eligibility and credit posting follow the in-app offer.
Artificial Analysis: GPT Image 2.6 Flare/Sunburst top AA image boards; Flare ~2× faster at same price
Boss lens: same price + faster gen raises creative throughput—pair leaderboards with your own blind tests (inference).
CollapseArtificial Analysis: GPT Image 2.6 Flare/Sunburst top AA image boards; Flare ~2× faster at same price
Per Artificial Analysis: GPT Image 2.6 takes #1/#2 on AA-Image-T2I and Editing v2.1. Flare is nearly 2× faster than 2.5 Flare with slightly better quality (~55s→~30s median); Sunburst is slower but leads. API pricing matches 2.5 (~$30/1M image output tokens; ~$0.21 per 1024² max-quality image). Third-party leaderboard—not a new OpenAI price sheet.
Crusoe: Google buys out SE Wyoming data-center site; project reportedly failed to lock customers incl. Google
Boss lens: AI campus ownership is still reshuffling—ask who finally holds land and power before long leases (inference).
CollapseCrusoe: Google buys out SE Wyoming data-center site; project reportedly failed to lock customers incl. Google
Per Techmeme citing Bloomberg: Crusoe says Google bought out its southeast Wyoming data-center site; the project reportedly failed to lock customers including Google. Other posts mention a prior ~$17B investment narrative—treat land buy vs. full campus spend carefully. Infra deal flash—terms/price not fully in the tweet.
geekbb recommends Herdr as a practical runtime for coding agents
Boss lens: agent runtimes are becoming a category—pilot before standardizing the stack (inference).
Collapsegeekbb recommends Herdr as a practical runtime for coding agents
Per @geekbb (Nov 6): recommends open-source Herdr (github.com/herdrdev/herdr) as “the runtime your coding agents live on,” saying it’s what they’ve been using lately. High bookmark engagement (~91). Practitioner tool tip—not a vendor earnings story; vet for production yourself.
Chuangxiang Yuedong raises ~¥60M seed with Skyworth strategic investment (LeiPhone)
Boss lens: appliance capital keeps betting on living-room AI hardware—watch if it diverts attention from your offline experience (inference).
CollapseChuangxiang Yuedong raises ~¥60M seed with Skyworth strategic investment (LeiPhone)
Per LeiPhone and follow-ons: Chuangxiang Yuedong closed a seed round of roughly RMB 60M+, with Skyworth strategic investment / Skyworth Digital among backers. Coverage links founders to Tencent Games alumni and an AI motion / living-room console bet. Early-stage financing—full cap table/valuation not fully public.
OpenAI partners with Hacktron to hunt vulnerabilities in its own systems, esp. third-party boundaries
Boss lens: even labs hire outside red teams for integration edges—put your SaaS seams on a recurring drill (inference).
CollapseOpenAI partners with Hacktron to hunt vulnerabilities in its own systems, esp. third-party boundaries
Per OpenAI (Nov 6): partnership with Hacktron to jointly pick high-priority areas for testing, focusing on how OpenAI systems interact with third-party services, plus expanded Daybreak Red access to frontier models. Relationship started via the bug-bounty program. Defensive collab note—no vulnerability count disclosed.
Cult.fit Codex case: 20+ internal apps by non-engineers; $100k+ software savings; support costs ~50%+
Boss lens: standardize the coding agent first, then coach functions—self-serve apps need a platform, not vibes (inference).
CollapseCult.fit Codex case: 20+ internal apps by non-engineers; $100k+ software savings; support costs ~50%+
Per OpenAI’s Cult.fit case study (dated Nov 5; recirculated in-window): the India fitness chain standardized on Codex and enabled non-engineers to build 20+ internal apps in recent months; one engineer rebuilt dashboard capability over a weekend, avoiding $100k+ annual software cost; customer-support costs fell 50%+. Company scale cited: ~708 centers and nearly 1M paid members. Vendor case-study figures.
lxfater: AI reverse-engineering is routine; cites GCSA Agent ~91.3% on CyberGym Level 1
Boss lens: vibe coding speeds shipping and attacking—add automated audit and dependency provenance to release checklists (inference).
Collapselxfater: AI reverse-engineering is routine; cites GCSA Agent ~91.3% on CyberGym Level 1
Per @lxfater (Nov 6): while builders ship fast with Claude Code/Codex, attackers also automate reverse engineering and exploit validation with AI. The post cites GCSA materials claiming ~91.3% success for GCSA Agent on CyberGym Level 1 and 380M+ ecosystem users, plus MCP supply-chain warnings. Success-rate/user figures are source claims—verify on GCSA originals.