AI Morning Brief · Oct 23, 2026
Top Stories
1.DeepSeek's ~$15B funding round moves toward closing, now with dual-source confirmation
Bloomberg broke the story Oct 22 and Reuters confirmed with four sources: the target grew from 50B yuan to ~100B yuan (~$15B) in two step-ups, with CATL and Tencent as the largest investors and new names including Midea, Meituan, Geely and GigaDevice plus Boyu, CPE, Legend and IDG — over 15 investors at a ~500B yuan pre-money valuation, expected to close within days; the story had previously been a single-source China Times report. Separately, Huxiu reports DeepSeek is preparing a STAR Market IPO (filing by year-end at the earliest) and has hired former Hillhouse VC partner Yan Wentao as CFO; Reuters notes the final figures are not yet settled (Oct 22 / Bloomberg via Reuters and TechCrunch)
Sources: techcrunch.com · bworldonline.com
2.Backfill: AI-discovered critical PHP vulnerabilities (CVE-2026-109618/619) patched in emergency release
CVSS 9.2, unauthenticated remote code execution, exposure reaching back to PHP 5.4 with ~1.3M internet-facing hosts affected; OpenAI's Oct 21 blog says Calif found the pair through its "Patch the Planet" program using GPT-5.6 Sol, while Wiz independently found them with Gemini 4 Argon and GPT-5.6 Sol; OpenAI calls it one of the most consequential security events in recent memory, pledged $1B in Daybreak credits, and coordinated with the White House Super Intelligence Force, ONCD and CISA; patches 8.5.12/8.4.27/8.3.36/8.2.35 were released a day early after details leaked. Missed by the two previous editions; dated Oct 21 US / Oct 22 Beijing, inside the window
Sources: pymnts.com · runtimewire.com
3.OpenAI's Decisions API goes generally available, powered by GPT-6 Luna
leaving public beta (limited preview at the Sept 29 DevDay, open beta from Oct 6): $0.10 per million input tokens, $0.01 per million cached input, output tokens free, up to 400 questions per request, ~150ms latency (about 10× faster than the Responses API); adds GPT-Live integration, prompt caching, image/JSON input, Zero Data Retention and HIPAA; Box, Datadog, Ramp, Sierra and Speak are already using it (Oct 22 / Unite.AI, Seeking Alpha relay)
Sources: unite.ai · tradingview.com
4.Follow-up: the ARTEX attack chain expands — ~20 Japanese firms targeted, Korean bank chain traced to "Autumn-27"
a major follow-up to the Storm-0940/ConnectWise disclosure: GMO IERAE's Oct 19 report found ARTEX traces in attack infrastructure targeting ~20 Japanese firms (9 in the same industry); WSJ says ~80 Japanese firms were hit; Yomiuri reports S&J's simulation extracted virtual-user data in 15 seconds by chaining ARTEX with Claude; CrowdStrike traced the Korean bank attack chain (Shinhan, KB, Hana, Busan, late Sept–early Oct) to "Autumn-27," which used DeepSeek v4.1-flash, GLM-5.3, Grok 4.6 and Claude Code (Oct 21–22 / eSecurity Planet, relayed)
Sources: esecurityplanet.com · techdefused.com
5.Follow-up: Anthropic S-1 could go public as early as Friday (Oct 23), IPO set for mid-November
Yahoo Finance (Oct 21, three sources): the S-1 may surface Oct 23 at the earliest; Reuters analysis (Oct 22): Anthropic is holding to a mid-November IPO with the roadshow starting the week of Nov 9, the lone survivor among tech IPOs (CoreWeave sold off, Elice fell 32% on debut, NVIDIA-linked Firmus Grid shelved its listing), with a potential $1.5T valuation (Oct 22 / PYMNTS, Reuters analysis relayed)
Sources: pymnts.com · northlandnewsradio.com
6.Micron's Taoyuan union begins a 24-hour warning strike — the first semiconductor strike in Taiwan's history
Nikkei Asia: the strike began 9 a.m. local Oct 22 with 1,994 votes (99% of voting members) authorizing it; the union demands a permanent 15% share of operating profit (vs. Samsung's 10.5% and SK Hynix's 10%); Micron's Taichung union already reached an agreement on Oct 21 (Reuters) (Oct 22 / Nikkei Asia via relays; an AI-memory profit-sharing dispute)
Sources: aiweekly.co · theledger.asia
7.NVIDIA and Samsung Electro-Mechanics sign a multi-year ABF substrate deal
SEMCO's Oct 21 official notice: the two will co-invest in NVIDIA-dedicated ABF substrate lines in Sejong and Busan under a multi-year agreement, with shipments starting in 2026; KED Global calls it SEMCO's largest FC-BGA contract ever (multiple billions of dollars, industry sources); echoes the 2.5× substrate-capacity expansion plan announced at GTC Berlin (Oct 21–22 / KED Global; a chip-supply-chain story, not a product fact)
Sources: kedglobal.com
8.xAI launches the SpaceX Super Intelligence Partner Network
the official Grok partner program with two tracks: Transformation Partners for enterprise rollouts and Marketplace Partners for plugins/integrations; 130+ companies are in early access and 30+ organizations are training engineers under the Forward Deployed Engineer certification; launch names include Bain, Databricks, NVIDIA, Snowflake, Booz Allen, EPAM, MongoDB and Twilio; applications at x.ai/partners; India's Tech Mahindra announced the same day it will bring Grok Build to enterprise software modernization (Oct 22 / Teslanewswire, CryptoBriefing)
Sources: teslanewswire.com · cryptobriefing.com
9.Anthropic releases Claude API version 2026-10-21, its first versioned API since 2023
the Messages API removes the temperature/top_p/top_k sampling parameters (sending documents on assistant turns or citations on user turns now returns 400 errors); SDKs and the ant CLI ship matching major versions; a breaking change developers must handle (reported Oct 22). Separately on Oct 22, the full Claude developer platform joined Google Cloud billing and IAM — API and agent tooling usable through Google Cloud credentials, billing and existing commitments, with Anthropic operating and processing the data (Oct 22 / PANews, RuntimeWire)
Sources: panews.io · runtimewire.com
10.Microsoft 365 Copilot becomes a standalone purchase from Dec 1; CSP channel defaults to usage-based billing
from Dec 1, 2026, Microsoft 365 Copilot can be bought separately from the base Microsoft 365 plan; new Copilot Business licenses via CSP default to usage-based billing (4,000 credits per person per month, $0.01 per credit ≈ $40/user/month). Copilot Code enters Frontier rolling release — natural-language app building/publishing with admin-controllable usage billing and optional local coding sessions (Oct 22 / Windows Forum, Message Center relay)
Sources: windowsforum.com · blog.tophhie.cloud
Platform & Model Tracking
OpenAI (GPT/ChatGPT)
Decisions API goes GA (see Top Story 3); usage-policy update (Oct 22): model routers/gateways need advance approval, effective Nov 20, and training competing models on model outputs is banned (TokenPost); Chan Park signed a letter to Governor Newsom (Oct 20) supporting a stronger SB53 — mandatory kill switch, annual independent audits, autonomous AI R&D inside the safety framework (TokenPost relaying The Information); GPT-6.1 "Astra" still unreleased as of Oct 23
Sources: tokenpost.com · tokenpost.com
Anthropic (Claude)
Claude API version 2026-10-21 (see Top Story 9); developer platform joins Google Cloud billing/IAM (see Top Story 9); the S-1/IPO timeline (see Top Story 5); Amodei's Oct 21 X essay — delaying broad Mythos-class model access by ~4 months for cybersecurity protection, accepting months of lost model lead (PivotNews relaying The New Stack); Claude science: 318,000 designed scorpion-toxin binders → 211 lab-tested → 17 hits, confirmed by two independent labs (PivotNews relaying the-decoder); flagship remains Claude Opus 5.5 (Sept 22)
Sources: pivotnews.ai · pivotnews.ai
Google DeepMind (Gemini)
Gemini on Apple CarPlay began rolling out Oct 22 (Gemini icon in the Maps interface, US and global rollout) and Personal Intelligence's Google Maps integration expands (Maps search and saved places, joint recommendations with Calendar) — minor feature launches; Google is notifying AI Ultra subscribers that Gemini 4 Argon will "soon" replace Deep Think (the 3.1 Pro reasoning tier); Argon processed nearly 25 trillion tokens in its first 72 hours with Pichai calling compute "super constrained" (Oct 21 Sources podcast); Gemini 4 Carbon still has no official progress as of Oct 23
Sources: 9to5google.com · techbriefly.com
Meta
Muse chat sharing was suspended over privacy concerns (Oct 22, WAM wire via UNN) — a product-function fact; at the Oct 21 earnings call Zuckerberg said the "Muse Charm" hardware device is "shipping soon" and announced a monetization tool for commercial transactions (single-source relay, unverified); no new Llama releases; Muse Wonder API pricing ($4/$20 per MTok, set Oct 12) unchanged; Wonder Full remains subscription-only; the new commercial terms reported earlier take effect Oct 30
Sources: unn.ae
xAI (Grok)
the SpaceX Super Intelligence Partner Network launch (see Top Story 8); Grok 5 still unreleased as of Oct 23 (grokmag's Oct 19 verification stands); the aibase rumor has no new evidence; no pricing or brand changes
Sources: teslanewswire.com
Microsoft
Microsoft 365 Copilot standalone purchase + CSP usage-based billing defaults (see Top Story 10); Copilot Code enters Frontier rolling release; no new Copilot/Azure AI model releases in window
Sources: windowsforum.com
NVIDIA
the official blog confirms Nemotron Bolt E4B (4GB local agent) releases Oct 23, but as of Oct 23 07:00 no launch announcement or weights were visible — treated as confirmed-plan/awaiting-landing, not a completed release; NeMo adds day-zero support for MiniMax M3.1 Flash, Liquid AI's LFM2.5-24B-A2B and Black Forest Labs' 5.6B open image model FLUX.3 Klein; the Samsung Electro-Mechanics ABF substrate deal (see Top Story 7); a claimed "global AI regulation is absurd" remark from Huang is single-source (huoxing via KuCoin) and unverified — see Unverified Items
Sources: blogs.nvidia.com
Mistral
backfill — the Voxtral 3 speech model family (Transcribe/TTS/real-time voice-agent API+SDK, expected $0.001/minute, 50+ languages) was officially announced Oct 20 and missed by the two previous editions; Large 4 weights not released as of Oct 23 07:00 (still promised by end of October)
Sources: pivotnews.ai
China AI Companies
DeepSeek
follow-up on the ~$15B (~100B yuan) raise closing with dual-source confirmation (Bloomberg + Reuters, see Top Story 1); Huxiu reports a STAR Market IPO is being prepared (filing by year-end at the earliest, aiming for a 2027 listing) and former Hillhouse VC partner Yan Wentao joined as CFO — Reuters notes the final figures are not settled and there is no official announcement yet; no new model releases (V4.1-Flash-1017 remains current)
Sources: techcrunch.com · huxiu.com
Moonshot AI (Kimi)
Kimi K3.1 is still NOT officially released as of Oct 23 07:00 — no announcement on the official blog or WeChat account; the only new signal is a ChainCatcher Oct 22 flash claiming "Kimi K3.1 and Qwen4 may launch this week" (single-source, unverified — see Unverified Items); Nomura's "before mid-October" prediction window has passed unrealized; K2.8 remains in preview
Sources: chaincatcher.com
MiniMax
M3.1 follow-up — correction: the official announcement date was Oct 19 (yesterday's edition said Oct 20); weights are promised for Oct 23 (official X: "Model weights will arrive on Oct 23") but no official confirmation of release was visible by 07:00 Oct 23; NVIDIA's blog already lists it as a day-zero-supported "new open model" with no delay notice; M3.1 Flash scores 43 on the AA Intelligence Index (MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Kimi K3 44)
Sources: the-decoder.com
ByteDance (Doubao)
five updates to Doubao Work (announced Oct 21, reported Oct 22) — local Office editing adds Word with direct local-file opening, a new traceable memory panel, plus cloud-PC, browser, message-queue and complex-task-division upgrades; Doubao PR chief Liu Xing denied the "team halved" rumor — the team was under 50 people, 11 affected with 3 departures, some roles moved to the Doubao trading/Work teams; CCTV's China Voice (Oct 22) investigated the AI-membership gray market (cracked tutorials, account rentals, API-key splitting on second-hand platforms) — Doubao said subscriptions only come through official channels and "unlimited freeloading" claims are impossible; Kimi said account renting/selling violates its terms and is being cracked down on
Sources: mrjjxw.com · m.dxhtj.com
Kuaishou (Kling)
Kling 4.0 launched on Adobe Firefly (Adobe's official X post, 13:07 Oct 22) and its API service debuted late Oct 21, opening gradually to enterprise customers and developers — specs: up to 30 seconds native generation, 10 keyframes, 4K 10-bit HDR, 21:9, 8,000-token prompt limit, 15 multimodal references; Kling claims 224 countries/regions, 100M+ users, 50K+ enterprise clients and ~$500M annual revenue scale (21st Century Business Herald, Oct 22)
Sources: ai123.com · 21jingji.com
Zhipu (GLM)
won the Xiamen Health-Medical Big Data Center's "medical LLM foundation and application engineering phase one" bid at 10.58M yuan (reported Oct 22); Zhidx counts at least 21 government/enterprise wins so far with 3 above 10M yuan; H1 2026 revenue ~954M yuan, up 399.7% year-on-year, with the open platform/API business at ~825M yuan (86.5%); a "GLM-5.3 API launched today" flash was verified as a rehash of the Aug 14 release — not credible, excluded
Sources: k.sina.cn
Tencent (Hunyuan/Yuanbao)
no new product facts in window (the Keep skill and Yuanbao Search technical follow-ups were reported Oct 21; the $5B note issuance is Oct 14 news; Tencent appears as a lead investor in DeepSeek's new round, see DeepSeek entry)
Alibaba (Qwen)
no model or product releases in window ("Qwen4 this week" is only the ChainCatcher single-source flash); market note: the Hang Seng Tech index fell over 4% on Oct 22 with Alibaba down over 4% to HK$98.95, below its Aug 26 HK$112.70 placement price (Yicai via Sina)
Sources: finance.sina.cn
Baidu
the Baidu Maps open-platform warned Oct 22 that cheap resale of its technical-service licenses (e.g. 3,000-yuan offers) is not official; Citi research expects Q3 results in mid-November with AI-cloud infrastructure revenue up 99% to 8.35B yuan and core ad revenue down ~21%, cutting the target price 7.2% to $154 (Buy maintained); Baidu World 2026 opened registration — Li Yanhong's keynote will unveil the latest "chip-cloud-model-agent" stack
Sources: finance.sina.cn · finance.sina.cn
Manus, StepFun, SenseTime
no significant in-window developments
Worth Reading
1."DeepSeek reportedly raises $15B days after US restrictions" — TechCrunch, 2026-10-22
DeepSeek's ~100B-yuan raise closing while US federal agencies move to remove it — funding details against the geopolitical backdrop.
Sources: techcrunch.com
2."AI-Found PHP Vulnerabilities Spark Emergency Patch" — AI Parabellum, 2026-10-21
a recap of the PHP2Shell incident — GPT-5.6 Sol and Gemini 4 Argon each independently finding the same critical pair, and the "defender window" concept.
Sources: aiparabellum.com
3."Analysis: Anthropic IPO stands alone as AI market jitters sidelined others" — Reuters (relayed), 2026-10-22
the full picture of Anthropic's lone-wolf run through an IPO winter (mid-November target, $1.5T potential valuation, contrasted with CoreWeave's selloff).
Sources: northlandnewsradio.com
Unverified Items
1.
RecursiveSI released Flint, an open MoE coding model (40B total / 2.5B active parameters, 262K context, claiming to beat Qwen3.6-35B-A3B on TerminalBench 2.1 and SWE-bench Pro v2): single-source (horizon.alchemylab.sh, Oct 22).
Sources: horizon.alchemylab.sh
2.
Jensen Huang calling global AI regulation "absurd," urging Nscale to push forward, and Firmus Grid abandoning a $30B IPO: single-source relay (huoxing via KuCoin), uncorroborated.
Sources: kucoin.com
3.
A ChainCatcher flash claiming "Kimi K3.1 and Qwen4 may launch this week": single-source, uncorroborated; neither has an official announcement as of Oct 23 07:00.
Sources: chaincatcher.com
4.
Whether NVIDIA Nemotron Bolt E4B (4GB) and MiniMax M3.1 Flash weights actually land on Oct 23: the blog confirms Bolt E4B for today and MiniMax promised Oct 23, but neither had landing evidence by Oct 23 07:00 — treated as confirmed-plan/awaiting-landing, not completed.
Sources: blogs.nvidia.com
5.
"Meta's Muse Charm hardware device shipping soon": single-source relay of the Oct 21 earnings call (PivotNews), uncorroborated.
6.
OpenAI's letter to Governor Newsom supporting a stronger SB53 (kill switch, annual independent audits): TokenPost relaying The Information; no official text seen.
7.
Items carried over with no in-window progress: Grok 5, Gemini 4 Carbon, Kimi K3.1, Zhipu's 3.5-trillion-parameter rumor, China's 23-AI-technology export controls (oodaloop single-source), OpenAI agents weakening monitoring (substack single-source), Claude tracing the NSA P-256 curve sequence (kimkj single-source); DeepSeek's final round figures remain unsettled (Reuters notes "details not finalized").