Vol. 02 · No. 05 — Monday, November 2, 2026
Covering Oct 26 — Nov 1, 2026
AI WEEKLY
The week in agents · coding · capital
The One-Line Theme
The Great
Repricing
Trillion-dollar checks get trimmed, filed, and wired to each other — while agent labor gets its first real price war.
In this issue — Agent labor repriced (P3) · The silicon beneath it (P4) · Copilot, the delivery system (P5) · The open-weight flank (P6) · Trillion-dollar fine print (P7) · The barbell market (P8)
Front Page
OpenAI cuts its raise to ~$20B at $1.25T
Financial Times, Oct 30
Also Inside
Anthropic files for a November IPO — ANT on Nasdaq
SEC filing via Yahoo Finance, Oct 26–31
And
Meta raises up to $95B — funded by its own chipmakers
CNBC · Yahoo Finance, Oct 26–27
Researched, written & produced end-to-end by the OnePress agent · getonepress.com/weekly/2026-11-02
Issue № 05
This Week at a Glance
Oct 26 — Nov 1, 2026
01
OpenAI trims the ask
Round target cut to ~$20B at a $1.25T valuation — down from $30B at $1.4T. First close may land below target.
FT · CoinDesk · Reuters, Oct 30
02
Anthropic goes public — on paper
IPO registration filed Oct 26 under “ANT” on Nasdaq; Q3 revenue $16.8B (+45% QoQ); trading could start as soon as Nov 16.
Yahoo Finance · Reuters · AI周报, Oct 15–31
03
Meta's $95B circle of capital
Meta lines up up to $95B for AI buildout; Nvidia commits $3B, AMD $1B — suppliers buying into their own customer.
CNBC, Oct 26–27
04
Copilot becomes a delivery system
GitHub Universe: Fusion ships GA, multi-repo Copilot projects, shared automations that learn — and Copilot bundled into GitHub Enterprise.
GitHub Changelog · Unite.AI, Oct 28
05
Agent labor gets cheaper
GPT-6.1 Terra: computer-use at $1/$5 per million tokens — Astra-level screen work at a tenth of the price (OpenAI claim).
OpenAI · MLQ News, Oct 26
06
Google TPUs leave home
Oracle AI World: TPUs arrive on Oracle Cloud in 2027 — the first hyperscaler outside Google to sell them.
Oracle · Data Center Knowledge, Oct 27
$20B
OpenAI's new target
$16.8B
Anthropic Q3 revenue
−5.3%
Nasdaq, worst month since Mar '25
$15.3B
Thrive–ADT take-private EV
50.3%
Terra on Terminal-Bench 4.0
Sources: Financial Times (Oct 30) · Yahoo Finance & Reuters (Oct 15–31) · CNBC (Oct 27) · GitHub (Oct 28) · OpenAI (Oct 26) · Oracle (Oct 27) · Yahoo Finance (Oct 30) · −5.3% is October's Nasdaq Composite monthly move, Yahoo Finance Oct 30.
AI Weekly № 05 · P2
Section A · AI Agents
Oct 26 — Nov 1, 2026
Agent labor gets a price cut — with the safety card stapled on
Lead · Model
GPT-6.1 Terra: screen work at half price
OpenAI's new mid-tier is built for computer use — reading screens, clicking through workflows, finishing repetitive tasks. Standard API pricing is $1 in / $5 out per million tokens: half of GPT-6.1 Sol, and 90% below GPT-6 Astra at list rates (derived from published prices).
OpenAI says Terra matches Astra on computer-use benchmarks at one-tenth the token price and finishes tasks 3× faster; Terminal-Bench 4.0 jumps to 50.3% from 21.5% for GPT-5.6 Terra. Independent check: Artificial Analysis scores it 48 (rank 25 of 221) — capable, not frontier, and priced accordingly.
Available now to Plus, Pro, Enterprise and Edu plans in ChatGPT Work and Codex; API model gpt-6.1-terra.
Sources: OpenAI announcement, Oct 26 · MLQ News, Oct 26
The Fine Print
Capability up, disclosure too
Terra's launch ships with an unusually explicit safety card disclosing agentic limits — part of a pattern this week:
Oct 30 — GPT-6 Rosalind Research ships with relaxed bio-refusals, gated behind trusted access only.
Oct 30 — ChatGPT restricts weapons guidance for teens and auto-ends chats showing violent fixation.
Oct 29 — Model Spec rewritten around five values; “the Assistant” rebranded as GPT.
Cheaper agents plus published boundaries is the week's real signal: the price war is arriving with receipts.
Sources: OpenAI, Oct 29–30 · Reuters, Oct 30 · MLQ News, Oct 26
Sources: OpenAI “Introducing GPT-6.1 Terra” (Oct 26) · MLQ News (Oct 26) · Reuters (Oct 30) · OpenAI release notes (Oct 29). Price comparisons derived from published list rates.
AI Weekly № 05 · P3
Section A · AI Agents
Oct 26 — Nov 1, 2026
The silicon beneath the agents — and an open-source counterweight
Infrastructure
Google TPUs head to Oracle Cloud
At Oracle AI World (Las Vegas, Oct 27), Oracle and Google announced 8th-gen TPUs — 8t for training, 8i for inference — on OCI starting calendar 2027: the first time Google's custom silicon is sold on a hyperscaler outside Google. Oracle calls it a “gigawatt-scale” bet.
The same stage added Cerebras wafer-scale inference (2027), SambaNova SN50, Positron Atlas (GA now) — and OpenAI's Broadcom-built “Jalapeño” accelerator. One cloud, five silicon stories, all aimed at agentic, long-context workloads.
Sources: Oracle newsroom, Oct 27 · Data Center Knowledge · Futurum, Oct 27–28
Open Source
DeepSeek open-sources an agent base
Days after its mega-round details landed, DeepSeek published an open-source agent infrastructure architecture (reported Oct 31) — the same week its ¥105.5B second round was confirmed with ~30 strategic investors.
The pairing matters: China's best-capitalized lab is giving away the scaffolding while Wall Street auctioneers the frontier. Watch whether Western labs answer with open agent runtimes of their own.
Sources: 量子位 (QbitAI), Oct 31 — single-source, flagged · eastmoney AI周报, Oct 31
Enterprise
Anthropic courts the enterprise — and the government
Oct 27: Anthropic expands its GSA partnership — a 40% discount on Claude Enterprise, Claude for Government and API for government partners (company announcement).
Oct 28: Claude Enterprise ships smart reports, in-app surveys and pooled group budgets; the developer platform lifts first-party API request limits from 32MB to 128MB — plumbing for document-heavy agents.
Sources: Anthropic newsroom, Oct 27–28 · Releasebot mirror, Oct 29
Reality Check
DoorDash says third-party AI agents have driven only “several hundred” orders against ~3 million deliveries a day. Agent commerce is a demo story before it is a P&L story. — AI Agent Store, Oct 31 (single source, flagged)
Sources: Oracle (Oct 27) · QbitAI (Oct 31, unverified second source) · Anthropic (Oct 27–28) · AI Agent Store (Oct 31, single source). Enterprise items are vendor announcements.
AI Weekly № 05 · P4
Section B · AI Coding
Oct 26 — Nov 1, 2026
GitHub Universe: Copilot stops being an autocomplete and becomes a delivery system
The Big Ship
Five releases, one direction: out of the editor
Copilot Fusion GA — the flagship experience leaves preview.
Copilot projects (public preview) — describe a goal, attach repos, mocks, Slack threads; get parallel agent sessions across multiple repositories, results delivered as reviewable PRs.
Shared Copilot Automations — team-owned cloud workflows, triggered by events or schedules, that now learn from outcomes and propose their own improvements for humans to approve.
Agentic Workflows GA + AI Scan for repositories enters public preview.
Sources: GitHub Changelog, Oct 28 · Unite.AI · Global AI Community, Oct 28
The Business Move
Copilot folds into GitHub Enterprise
From Oct 28, GitHub Enterprise customers get every Copilot feature — app, code review, automations — with no separate subscription; they pay usage-based AI credits instead. Existing Business/Enterprise plans can stay or migrate.
Alongside it: a connected Copilot cloud experience — sessions persist across web, mobile, CLI and IDE, running in secure cloud sandboxes with resumable snapshots. Sandbox billing begins November 2026.
The read: GitHub is pricing agents like compute, not seats — and betting that automations, not chat, are where enterprise budgets land.
Sources: GitHub Changelog, Oct 28 · Unite.AI, Oct 28
Sources: github.blog changelog entries (Oct 28) · Unite.AI GitHub Universe recap (Oct 28) · globalai.community Day-1 roundup (Oct 28).
AI Weekly № 05 · P5
Section B · AI Coding
Oct 26 — Nov 1, 2026
The open-weight flank attacks from below
Edge Models
Liquid AI's LFM3: a 2B-class family that beats a 9B
On Oct 27, Liquid AI shipped six open-weight models — text, vision, audio, omni, plus two decision models — branched from one 1.7B tri-hybrid base trained on 22T tokens.
The text model averages 49.37 across its reported suite vs 47.24 for Qwen3.5-9B at roughly one-fifth the parameters, with a 256K context and a ~1GB footprint at 4-bit. Agent scores lead the class: WebArena-Verified 38.42 vs 3.02 for Liquid's previous generation.
The thesis: most agent deployments don't need the frontier — they need a gigabyte and a harness.
Sources: Local AI Zone tracker, Oct 27 · Hugging Face trending, Oct 27–29
Also Shipped
Weights drop and data-plane plumbing
Mistral Large 4 weights released (Oct 29–30) — the trillion-parameter “Le Chonk” preview graduates to published weights, pitched “built for coders, trusted by defenders.”
Snowflake Cortex coding agents gain public controls (Oct 28): a thinking_effort dial, system-instruction injection, and versioned stage_mounts into sandboxed runtimes.
ChatGPT release notes (Oct 29): more Work and Codex usage carved into Pro and Business Premium plans.
Pattern watch: the frontier's top rung stayed quiet this week — the action was in the middle of the price ladder and at the edge.
Sources: Mistral, Oct 29 · hellogrowth tracker, Oct 30 · Snowflake docs, Oct 28 · OpenAI release notes, Oct 29
Sources: Local AI Zone (Oct 27) · Mistral “Le Chonk” (Oct 29) · Snowflake documentation (Oct 28) · OpenAI help center release notes (Oct 29). Benchmark figures are vendor-reported unless attributed.
AI Weekly № 05 · P6
Section C · Capital
Oct 26 — Nov 1, 2026
Trillion-dollar fine print
Who
What happened
The number
OpenAI
Trimmed its raise to ~$20B at a $1.25T valuation (FT) — down from $30B at $1.4T in September. Friday deadline passed with the first close expected below target; UAE-led consortium continues. Revenue run-rate reportedly near $70B by year-end (unaudited).
−$10B target cut
−$150B valuation cut
FT · CoinDesk · Bloomberg, Oct 22–30
Anthropic
Publicly filed its IPO registration on Oct 26 — ticker “ANT” on Nasdaq, Goldman Sachs & Morgan Stanley leading. Q3 revenue $16.8B, +45% QoQ (vs +147% in Q2); adjusted-profitable and cash-flow positive at $1.70 revenue per $1 of compute. Debut possible as soon as Nov 16; valuation talk spans $1.5–2T, with some investors tempering on the growth deceleration.
$16.8B Q3 revenue
+45% QoQ
Yahoo Finance · Reuters · CNBC, Oct 15–31
Meta
Announced plans to raise up to $95B (Oct 26): $35B underwritten first, up to $60B via at-the-market program — its first negative free-cash-flow quarter in the background. Then the twist: Nvidia committed $3B, AMD $1B (derived: combined $4B), with Broadcom eyeing up to $5B aggregate. Suppliers are now financing their own customer.
$95B raise plan
$4B from chipmakers
CNBC · Yahoo Finance, Oct 26–27
DeepSeek
Second-round details confirmed Oct 30: ¥105.5B (~$15B) from ~30 investors led by Tencent and CATL; post-money reported at ¥615.5B (~$86B). Founder Liang Wenfeng says ARR has passed $1B; Shanghai IPO prep reportedly targets 2027. Cumulative raise now exceeds ¥150B in six months.
¥105.5B (~$15B)
~30 investors
Bloomberg via Investing.com · eastmoney AI周报, Oct 30–31
Sources: Financial Times via CoinDesk/AI Weekly (Oct 30) · Bloomberg (Oct 22) · Yahoo Finance & Reuters (Oct 15–31) · CNBC (Oct 26–27) · Bloomberg via Investing.com & eastmoney (Oct 30–31). ¥ figures converted at ~7.15 CNY/USD — approximate. Reported figures, unaudited unless filed.
AI Weekly № 05 · P7
Section C · Capital
Oct 26 — Nov 1, 2026
The week's checks, drawn to scale
Anthropic IPO (target talk)
up to $100B
Meta equity program
$95B
OpenAI round (new target)
~$20B
DeepSeek Series II
~$15B
Thrive → ADT (take-private EV)
$15.3B
Upwind Series C
$400M
Serval Series C
$150M
Linear scale — the mid-tier bars are nearly invisible by design. That is the story: a barbell market.
Markets Flinch
Nasdaq Composite just logged its worst month since March 2025 (−5.3% in October); Reuters counts >$2T wiped from the Nasdaq-100. Friday closed −2.9% on Amazon/Intel capex scare lines. J.P. Morgan still models ~$900B of AI capex in 2026 → $5.5T by 2030.
Sources: Reuters, Oct 30 · Yahoo Finance · CNN, Oct 30
Mid-Tier Bench
Upwind: $400M Series C at $3.8B (Bessemer-led; revenue 4× in 2026, company claim). Serval: $150M at $3B (Sequoia-led; third round in 12 months, pitched as the ServiceNow replacement). Thrive–ADT: Josh Kushner's OpenAI-linked holding takes the 150-year-old alarm company private at $10/share; ADT jumped ~33%.
Sources: GlobeNewswire, Oct 28 · Business Wire/Reuters, Oct 29 · TechCrunch · WSJ, Oct 29
Sources: Yahoo Finance (Oct 26), CNBC (Oct 27), FT (Oct 30), Bloomberg via Investing.com (Oct 30), GlobeNewswire (Oct 28), Business Wire & Reuters (Oct 29), TechCrunch/WSJ (Oct 29), Reuters & Yahoo Finance (Oct 30). “Up to $100B” Anthropic target is press-reported talk, not a filed term.
AI Weekly № 05 · P8
Worth Noting · Policy, Safety, Open Web
Oct 26 — Nov 1, 2026
The quiet stories that will outlast the fundraises
Standards
One robots.txt line to opt out of AI training
An IETF draft by Anthropic's Kevin Kelley (published Fri, Oct 30) defines a shared policy token — __usage_ai-training — that lets any site refuse generative-AI training with a single robots.txt group instead of four crawler-specific blocks. Google, Microsoft and OpenAI announced support the same day; Google says it squares with the EU GPAI Code of Practice. Publishers from the FT to the BBC were consulted.
Four rivals agreeing on one signal is rare. Enforcement is the open question — last year, both OpenAI and Anthropic were accused of skirting robots.txt.
Sources: The Next Web, Oct 30 · The Verge · Washington Post, Oct 30
Safety Ledger
Capability with attached guardrails
Rosalind, gated. OpenAI shipped GPT-6 Rosalind Research with relaxed bio-refusals — locked behind trusted-access vetting, the Argon playbook.
Teens first. ChatGPT restricts weapons guidance for minors and auto-ends chats exhibiting violent fixation (Reuters, Oct 30).
SDK hygiene. OpenAI Python SDK 3.42.0 adds agent context compaction, session review and org key management (Oct 31).
Sources: OpenAI, Oct 30–31 · Reuters, Oct 30 · GitHub, Oct 31
Consumer AI Goes East
Apple confirmed Apple Intelligence lands in China starting Nov 4 (developer beta; GA in December with iOS 27.3) — powered with Alibaba, using Qwen for text and image understanding. The first mainstream US–China model partnership inside a consumer OS. — eastmoney AI周报, Oct 31 (reported)
Sources: The Next Web (Oct 30) · The Verge & Washington Post (Oct 30) · OpenAI (Oct 30–31) · Reuters (Oct 30) · eastmoney AI周报 (Oct 31). Robots.txt enforcement claims reference Business Insider's 2024–25 reporting.
AI Weekly № 05 · P9
Outlook · Next Week & Beyond
From the Nov 2, 2026 issue
What to watch
Nov 3 — 7
The Anthropic roadshow. Institutional marketing begins in earnest; filing math points to a possible Nov 16 first trade. Watch the priced range against the $1.5–2T talk — it is the market's referendum on frontier-AI margins.
OpenAI's first close. The FT says it may land below the new $20B target. Any shortfall is the loudest possible data point on private-market appetite.
Nov 4
Kimi K3.1 Pro full weights drop — the highest-ranked open-weight model ever measured (AA index 55) goes fully public. Apple Intelligence opens in China the same day.
Later in November
Copilot cloud sandbox billing starts. GitHub's usage-based agent pricing gets its first real demand curve — the test of “agents as compute, not seats.”
Earnings echo. Amazon, Intel and Alphabet's capex lines set the tone; every hyperscaler guidance revision now moves the whole AI complex.
The robots.txt test. First major publisher to flip the token — and first crawler accused of ignoring it.
One-Line Takeaway
The checks got smaller, the disclosures got bigger — and for the first time this cycle, capital is pricing AI like a business instead of a bet.
Forward-looking items are expectations, not facts: Anthropic timing per Yahoo Finance/Reuters (late Oct) · OpenAI close per FT (Oct 30) · Kimi weights per Local AI Zone (Oct 23) · Apple China per eastmoney AI周报 (Oct 31) · Copilot billing per GitHub Changelog (Oct 28).
AI Weekly № 05 · P10