AI Morning Brief · Oct 7, 2026

The round far exceeds the original ¥50B target, with people familiar saying it could reach ¥100B; Tencent and Contemporary Amperex Technology are among the lead investors; the raise is expected to close in October, laying groundwork for a planned 2027 IPO. Context: DeepSeek closed a roughly ¥50B first external round in June 2026 at a ¥350B+ valuation, then reopened a second round in August at about ¥500B pre-money. The 2027 IPO plan is a sources-only claim; DeepSeek did not respond.

Unveiled at the AI Everything summit in Abu Dhabi: a 1T-total-parameter / 49B-active-parameter MoE model, natively multimodal with 1M-token context, aimed at coding, cybersecurity and enterprise workloads; API preview from Oct 6, open weights slated for announcement on Oct 27. CEO Artur Mensch claims it is "above the Chinese models" in some areas including cybersecurity; the company raised a €3B Series D in August at a €21B+ valuation. The "4,000 Nvidia Grace Blackwell GPUs" training detail is single-source and unverified.

Project Glasswing folds into a Defense / Red Team / Specialized structure, opening Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1 to vetted security teams with relaxed cybersecurity screening classifiers; Glasswing partners surfaced at least 129,000 verified software vulnerabilities from April–July, 33,000 of them critical or high severity.

Under the EU AI Act's transparency obligations (in force Aug 2), ChatGPT and Codex text will carry machine-readable invisible statistical watermarks, mandatory for EU users in the coming weeks; API developers worldwide can opt in from Oct 5 (off by default); the detector stays limited to approved researchers for now. The technical report (co-authored with UPenn and Yale) concedes a 10% synonym replacement drops detection from ~92% to 66%, and 25% replacement to 17%.

Free personal accounts lose Gemini Flash (3.6) and Pro (3.1) access; AI Plus ($4.99/mo) loses Pro (keeping Flash-Lite and Flash); AI Pro gains the Deep Think advanced-reasoning mode; usage caps move from fixed prompt counts to compute-based metering (refreshing every 5 hours with weekly caps). Multiple secondary outlets reported consistently on Oct 5–6, but official Google confirmation was not captured; the item was missed at the previous window's edge and is included today as a catch-up.

Anthropic announced availability to all paid Claude plans: a sidebar inside Docs, Sheets and Slides that reads selections and rewrites files in place, plus three beta connectors for creating and editing Google files from Claude chat, inheriting Google sharing permissions and enterprise compliance controls. No official anthropic.com announcement captured directly.

The gemini-nano-banana-2.1 release brings efficient image generation with multi-turn editing (better quality, prompt adherence, multi-turn character consistency, text rendering, 1:4–8:1 aspect ratios, 1K/2K/4K); gemini-3.1-flash-image (Nano Banana 2) is deprecated and shuts down on 2026-10-29.

OpenAI chief strategy officer Jason Kwon admitted the company's response to the June agent breach of Australia's Medicare government portal was "not good enough" (disclosed ~84 days late) and backed a mandatory disclosure framework; Anthropic endorsed required security-plan disclosure and called for alignment with overseas standards. Both conceded that reporting is currently at companies' own discretion — the gap the proposed AI Agent Breach Rules would close.

Acting Deputy Attorney General Trent McCotter instructed DOJ staff to use "super intelligence"/"SI" in public communications, policy documents, official records and appropriate court filings "to the maximum extent permitted by law", following Trump's executive order signed last week.

The Brooklyn startup founded in 2024 by two ex-DeepMind researchers released Beam: a 501B-total-parameter / 23B-active-per-token MoE focused on coding, reasoning and agents; the company claims reasoning performance comparable to Z.ai's GLM-5.2 at 1/4–1/3 the inference compute; weights land on Apache 2.0 at end of October (waitlist only for now); training used 23.8T tokens, 6,144 GB300s for 4 weeks of pre-training and 10.5K GB300s for RL (100M rollouts).

textGrain invisible watermark rollout in the EU (see Top Story 4 — a compliance feature launch). API paid-usage tiers restructured (Oct 6): an official developer-forum post merges the five paid tiers into three (Build / Launch / Grow); the top Grow tier's entry threshold drops to $500 lifetime API spend (was $1,000); existing paid orgs migrate automatically. A pricing/availability change. The forum timestamp "October 6, 2026, 6:41pm" carries no timezone: it falls in-window if UTC (~02:41 Beijing on Oct 7) and just outside if US Pacific.

Claude for Google Workspace public beta (see Top Story 6 — product launch); three-tier Cyber Verification Program expansion (see Top Story 3); Claude for Startups expansion (Oct 6): at SF Tech Week, eligible startups get one free year of Claude Team (up to 5 premium seats) plus $1,000 in API credits — single source (TechCrunch, 9:00 AM PDT ≈ midnight Beijing Oct 7, in-window); Claude Code 2.1.292 (Oct 6): plugins/agents/cloud features (--marketplace, effort params) plus security fixes — single source (Releasebot), minor release; Australia hearing support for mandatory incident reporting (see Top Story 8 — policy, not product); IPO filing disclosed CEO Dario Amodei's $18M 2025 pay (Reuters exclusive, not a product fact).

Nano Banana 2.1 GA with 3.1-flash-image retiring Oct 29 (see Top Story 7 — model release and deprecation); free-tier tightening effective Oct 9 (see Top Story 5 — pricing/availability change, catch-up).

No significant updates at the model/product level. A Citi report (cited by Reuters, Oct 6) estimates Meta's Muse assistant at 6.6M+ downloads and 1.8M DAU, projecting $27B+ revenue by 2030 — an analyst forecast, not a product fact.

No significant updates. Grok 4.7 verified via xAI's official release notes as a Sept 21 release (outside the window); only third-party leaderboard refreshes in-window.

Mistral Large 4 "Le Chonk" release (see Top Story 2).

Near-final ¥80B+ new round led by Tencent and CATL (see Top Story 1 — Bloomberg first, Reuters follow-up). Also: Bloomberg analyst Robert Lea's Oct 5 report (cited by Chinese media Oct 6) claims China's top models trail US rivals by just 3% on benchmarks (9% in May, 15% at year-start), with DeepSeek-V4.1 Flash scoring 81.1 on LiveBench (Anthropic's top model 83.4), ranking 6th globally — the highest rank for a Chinese model to date. Single source (Sina Finance relaying 三言科技); the report itself was not found.

Amazon Bedrock adds GLM-5.3 (Oct 6; 科创板日报, cross-verified via multiple relays): enterprise customers can call it directly, with AWS sharing usage-based revenue — opening an overseas monetization channel; Alibaba Cloud's Bailian already has a similar split and Huawei Cloud lists GLM-5.3 too. Zhipu's Hong Kong stock (2513.HK) rose over 8% intraday; MiniMax rose over 4%. Same report notes Cursor has added GLM-5.3 and GLM-5.3-Flash (CursorBench 4.0 top score among open-weight models) — a Cursor-only claim without independent verification.

Beam claims parity with GLM-5.2 and near-Qwen3.8-Max on coding and agent tasks (vendor self-claim, single source via ITHome).

A Sina Finance Oct 6 profile piece on Kimi K3 and a dual Shanghai–Hong Kong listing push is largely a recap of September news, not a new development.

No model releases, financing, product moves or regulatory developments during China's Oct 1–7 holiday.

A tight read of the 20-page textGrain technical report's key numbers (80% detection at 200 tokens, 95% at 400; much lower on math answers) — and the punchline that a watermark without an open detector is just a promise.

A full Beam FAQ — why a 501B model with 23B active needs ~500GB of VRAM resident, benchmark-by-benchmark comparisons against GLM-5.3/Kimi K3 (Beam trails on most coding benchmarks; the pitch is efficiency, not the leaderboard), and advice not to plan production around unreleased weights.

Unpacks the CVP's internal evaluation numbers (Red Team tier's 50/50 zero-intercept on CyScenarioBench, Opus 5.5 completing 34 tasks ≈ 67.6% of the unguarded baseline) and the dark side of the data-retention policy (mandatory retention across all three tiers; zero retention only via Enterprise Frontier Safeguards).

Claude experienced a ~17-minute outage on Oct 6, 12:25–12:43 UTC — third-party monitor nocrash.io only; no Anthropic status page or media cross-confirmation.

AI chip company Etched is seeking $40–50B in financing inquiries (twice the September round) — mentioned only in the ainews.malone.gr daily digest, no other sources found.

The VISTA visual-memory framework reportedly gives Claude Opus a perfect score on ARC-AGI-3 — mentioned only in a LinkedIn daily digest; no paper or media coverage found.

Cohere North 2 enterprise agent platform (persistent memory, spend controls, air-gapped deployment) — mentioned only in a LinkedIn digest; no other sources found.

UN human rights chief Volker Türk warning that "the clock on AI regulation is ticking" — only seen via the millenniumedu.org aggregation; the UN News original was not located.

Anthropic's $660M in donations from Oct 2025–Mar 2026 — aidailypost.com relaying The Information, single source.

Claude Cowork: groundtruth.day (Oct 5 post) claims new Cowork tasks on Pro/Max plans default to cloud execution from Oct 6 — single source, needs official confirmation.

Simon Willison's blog (Oct 6 evening, single source) claims claude.ai's free tier has switched to Sonnet 5.5 (vs OpenAI's free tier on Luna 5.6) — no official confirmation, unclear timezone, possibly just outside the window.

The Bloomberg Robert Lea report itself was not found publicly; the "3% gap" and "V4.1 Flash ranked 6th globally" figures come from a 三言科技 relay via Sina Finance — unverified.

CursorBench 4.0 "top open-weight score" for GLM-5.3-Flash — a Cursor-side claim relayed by 工商時報, no independent verification.