EAIDaily — October 12, 2026 (Issue #36)

English edition · Focus: AI Coding + Embodied Intelligence · Sources: AI HOT API + WebSearch triangulation

🧠 AI Coding

1. GPT-6.1 Sol default context jumps 272k → 500k in Codex & ChatGPT Work

What: OpenAI’s Tibo (Thibault Sottiaux) announced that GPT-6.1 Sol’s default context in Codex and ChatGPT Work rises ~80% to 500k tokens (max 1M, user-overridable), reducing compaction frequency for long-horizon tasks. This landed as Day 6 of the 28-day streak, paired with a global usage-limit reset confirmed at 04:01 UTC Oct 11 for eligible Codex/ChatGPT Work users. Why it matters: Compaction is where agent context gets lossy and where long coding tasks fail. Making 500k the default (not an opt-in) signals that inference-time configuration is now a headline product lever — and that OpenAI is competing on “fewer context resets” as a first-class feature. Source: AI HOT / x.com/thsottiaux; aiindustrytoday.com

2. Meta & Microsoft cut internal Claude Code usage; Google pushes Gemini workplace agents + “Carbon”

What: Per 商傳媒 weekly review (Oct 12), Meta and Microsoft are sharply reducing employee use of Anthropic’s Claude Code in favor of in-house tools — Meta’s Muse Code and Microsoft’s Copilot for Business. Google simultaneously launched Gemini workplace agents for enterprise coding/content tasks and is internally testing a Gemini 4 model, codenamed Carbon, reported to rival Anthropic’s top model. Why it matters: Anthropic’s biggest revenue risk isn’t OpenAI — it’s its own largest customers becoming competitors. Hyperscaler in-housing marks the end of the “everyone rents the frontier lab’s agent” era and a structural shift toward vertically integrated coding-agent stacks. Source: n.yam.com weekly industry digest

3. Cockroach Labs runs coding agents like a teaching hospital — 5 months of data

What: Cockroach Labs published numbers from five months of treating GitHub issues as “patients” and Claude agents as a medical team: triage, specialist agents, and human attendings signing off before code reaches production. Why it matters: Most agent-ops content is anecdote; this is one of the first longitudinal, org-scale datasets of a multi-agent maintenance workflow. The “teaching hospital” framing (supervised autonomy with escalating trust) is emerging as the credible enterprise pattern — contrast with blind full-autonomy demos. Source: ai-tldr.dev digest, Oct 12

4. Anthropic ships Claude Code 2.1.296: pinned workflow-agent models + earlier subagent compaction

What: Claude Code now lets you pin workflow agents to a single model and gives each subagent its own auto-compact schedule (autoCompactWindow), alongside gateway/reliability fixes. Meanwhile the Claude Managed Agents platform gained dynamic workflows (multi-phase background workflows, Oct 9 beta). Why it matters: Fleet operators keep asking for cost control and determinism per role — one model for the orchestrator, a cheaper one for workers. These are exactly the knobs that turn “agent demo” into “agent payroll”: fleet economics are becoming a product surface. Source: ai-tldr.dev; releasebot.io

5. Deno joins Cloudflare; Deno Deploy shuts down in six months

What: The Deno team is being absorbed by Cloudflare; Deno Deploy users have ~6 months to migrate. Why it matters: Another independent JS runtime/edge platform folds into hyperscaler infrastructure. For agent-era deployment (ephemeral per-task runtimes), consolidation means the “deploy a whole app from an agent” pipeline increasingly runs through a handful of providers — convenience up, optionality down. Source: ai-tldr.dev (ECOSYSTEM: MAJOR)

6. Lean proofs need more scrutiny as AI writes them at scale — Thomas Hales

What: Thomas Hales (Kepler-conjecture prover) warns that a Lean proof is only as trustworthy as Lean’s kernel and toolchain itself, and AI systems are now producing millions of lines of formal proof code, with human review lagging. Why it matters: “Formally verified” has been marketed as the answer to LLM hallucination in math, but the verification substrate becomes the new trust bottleneck — a verification recursion. Relevant to any coding agent workflow that treats compiler/typechecker pass as “done.” Source: ai-tldr.dev digest

🤖 Embodied Intelligence

7. Frost & Sullivan: Chinese firms sweep global humanoid revenue top-5

What: The 2026 Global Humanoid Robot Market Report (released Oct 12 coverage) puts 2026 H1 global shipments at ~36,000 units and revenue at $540M. By revenue, the top five are all Chinese: UBTECH (¥620M, 16.5%), Agibot/Zhiyuan (¥610M, 16.1%), Unitree (¥580M), Galbot, Leju. In the full-size embodied-intelligent segment (~17% of volume), Chinese vendors again take all top-5 revenue spots — that segment delivers ~60% of global revenue despite being a minority of units. Why it matters: Two signals: (1) China’s lead is now a revenue lead, not just a shipment-count story — the premium segment is also Chinese-dominated; (2) customers pay roughly 3.5× more per unit for genuine embodied intelligence vs. pre-programmed robots, validating “intelligence density” as the pricing axis. Source: 21财经 / Frost & Sullivan report via so.html5.qq.com

8. Unitree hits 5,000 G2 Pro humanoids produced; CIIF Shanghai opens first humanoid zone

What: Unitree announced (Oct 12) its 5,000th G2 Pro humanoid off the line, deployed across Chinese manufacturing and logistics. The same day, the 26th China International Industry Fair opened in Shanghai (300k㎡, 3,000+ exhibitors) with its first dedicated humanoid robot zone: Fourier’s GR-3C “astronaut” (55 DoF, teleoperation) debuted, alongside SIASUN, Tihoo, Unitree demos. Separately, Shenzhen customs data show robot exports up 10.5× YoY (Jan–Jul, ¥8.58B). Why it matters: Serial-production milestones (Unitree 5,000; UBTECH 10-min/unit factory; ZQ Robotics 15-min/unit) show China has moved humanoid robotics from lab to manufacturing discipline, while a national expo giving humanoids a dedicated hall institutionalizes the category. Export growth at 10× means the volume is leaving the country, not just being absorbed domestically. Source: humanoidapplications.com calendar; expo-ciif.org.cn; 南方日报

📡 Radar (watch, not yet confirmed/standalone-worthy)

Codex desktop steering & banked resets — faster mid-run correction + Queue-for-next-turn; control-plane polish, not new models (aicrier.com).
Claude Code Projects backlash day-1 — ignores global skills/CLAUDE.md, defaults to Opus high-effort; cache-TTL debates (1h vs losing ~15% weekly limit). Fleet-cost UX is the new battleground.
Mistral Large 4 open weights expected by end of October — “sovereign AI” option for enterprises.
Talent war: OpenAI hired ex-Cursor EMEA lead Ismail Elmas after halting model supply to Cursor; OpenAI/Anthropic poaching into India/Singapore.
Windows Copilot “PC insights” experimental feature — conversational hardware diagnostics; opt-in, US-only; credibility still AI-hallucination-limited.
DeepSeek “text-as-images” memory (Two Minute Papers Oct 12) — storing long text as images to save tokens; if verified, a novel token-economy direction.
xEI vs Minnesota AI-content obscenity ban appeal: oral submissions Oct 29; outcome could shape AI-generated content law.

🔑 Key Trends

Customers become competitors. Meta/Microsoft in-housing Claude Code shows frontier labs’ enterprise revenue has a structural ceiling: your biggest buyers will absorb the capability. Differentiation must move up the stack (Managed Agents, Cyber Mission) faster than in-housing can.
Inference-time configuration is the new shipping cadence. Default context 500k, per-subagent compaction, model pinning, steering latency — the week’s wins are almost entirely controls, not raw capability. Fleet economics IS the product.
Supervised autonomy is winning the enterprise narrative (Cockroach Labs teaching hospital; Hales on Lean verification) — trust architecture, not autonomy demos, is what gets published and adopted.
Embodied AI’s scoreboard went financial. With Chinese vendors holding both volume and premium-segment revenue top-5, the industry’s phase transition is from “can it walk” to “who books the revenue” — and China is booking it at 10× export growth.

Quality Check

Excluded as repeats of Issues #32–35 (Oct 7–11): Claude Code Projects launch, OpenAI 28-day streak days 1–5, Anthropic misalignment reports, Figure 03, Agility Digit #5,000, Google Gemini universal agent, Microsoft-Decision-1, State of AI Report.
GPT-6.1 Sol 500k is a new change (Day 6) distinct from prior-day streak items.
Frost & Sullivan report is fresh Oct 12 coverage; note it partially overlaps IDC H1 data cited Oct 6 (both ~25k/36k-class H1 figures from different methodologies — Frost & Sullivan reports 36k units H1).
The “humanoidapplications.com calendar” entries for Oct 12 (Unitree 5,000; Sanctuary AI/Flex) partially aggregator-generated — Unitree milestone corroborated by its placement in Frost & Sullivan ecosystem reporting; Sanctuary/Flex treated as unconfirmed (radar-worthy only).
Calendar events dated after Oct 12 and marked “Rumored” were excluded entirely.

Sources: AI HOT API (aihot.news/api/v1) + 4 WebSearch passes (商傳媒 weekly, ai-tldr.dev, releasebot.io, 21财经/南方日报, expo-ciif.org.cn, aiindustrytoday.com, aicrier.com).