EAIDaily — 2026-09-22

English AI Daily Brief | Focus: AI Coding & Embodied Intelligence Automation series, Issue #17 · Coverage window: Sep 21–22, 2026

Today’s Master Frame

Same price, more machine — and the boom gets its first honest census. The capability-per-dollar frontier collapsed on two fronts in 24 hours: xAI’s Grok 4.7 ships near-frontier coding at an unchanged $2/$6 per million tokens, and Xiaomi open-sources a 1.02-trillion-parameter MIT-licensed model that scores as the strongest open-weights system ever measured — at 1/20 to 1/60 of closed rivals’ prices. Meanwhile the embodied-AI boom met its first hard baseline: the IFR’s first-ever humanoid census counts ~7,000 units sold worldwide in 2025, against 2030 forecasts that now run to 890K — a 100× gap that is the investment story. The day’s second thread is trust infrastructure: Zhipu closed its ZCode data scandal with full open-source plus third-party audits, Kimi Code Desktop turns agent authorization into a product layer, and a 5-million-view “enter-key engineer” post put the human cost of agentic coding on the record. Compute, meanwhile, keeps marching from the datacenter to the factory floor (Hygon’s “physical world” silicon, unveiled today).

1. Grok 4.7 launches: frontier-adjacent coding at exactly Grok 4.6’s price

What happened. xAI released Grok 4.7 (Sep 22), positioning it as its strongest coding-and-knowledge-work model: a larger base model than 4.6, extended RL on multi-hour tasks, and training focused on self-verification and long-context management — plus native familiarity with the Grok Bot harness. Benchmarks moved most where agents live: Terminal-Bench 4.0 20.3% → 38.0% (near-doubling; edges GPT-5.6 Sol Max’s 37.3%), CursorBench 4.0 40.4% → 46.3% (above Sol Max’s 41.7%, behind Fable 5.1 Max’s 51.8%), DeepSWE v1.1 71.0%, and on Artificial Analysis the Coding Agent Index jumped 47 → 56, ranking 4th among native harnesses behind Fable 5.1, GPT-6 Astra and Opus 5. It leads outright on Harvey Legal Agent (19.6% vs Sol’s 2.5%) and EEBench (64.0%). Specs: 500K context (cutoff May 2026), Low/Medium/High/xHigh reasoning, a Fast variant (2× speed, 2× price), and a rebuilt safety stack (3.3% high-risk pass rate on HackerBench v0.3; LatchBio biosafety 62.4%). Pricing is unchanged: $2/M input, $6/M output (≤200K prompts; $4/$12 beyond; cached $0.50). Day-one availability: xAI API, Grok Build (default), Cursor (all paid plans), Devin, OpenRouter/Vercel/Cloudflare. Caveats: xHigh burns ~81K output tokens/task (vs 38K for 4.6), averaging $8.82/task in AA coding evals, and community testing found it worse than 4.6 on some exploit-crafting setups and weak at one-shot frontend generation versus DeepSeek V4.1-Flash. Context: Financial media report OpenAI is reworking ChatGPT/Codex agent capabilities into a packaged consumer agent to counter xAI’s GrokBot and Meta’s Muse.

Why it matters. The competitive frontier is no longer “who scores highest” but who closes dev loops per dollar — Grok 4.7 buys ~90% of Fable 5.1’s terminal work at ~1/5–1/8 the unit price, and the same price as its predecessor means the floor keeps dropping without a price war being declared. Its harness-native RL training (optimized for Grok Build’s short-step, feedback-rich loops — and measurably better there than in external harnesses) is further evidence that model and harness are now co-designed products, not separable layers. The OpenAI repackaging report confirms coding agents have become the consumer battleground of the second half.

2. Xiaomi open-sources MiMo-V2.6: the strongest open model on record, MIT-licensed, trained live in public

What happened. Xiaomi released and fully open-sourced MiMo-V2.6 (Sep 21–22): Pro at 1.02T total / 42B active parameters (sparse MoE, 384 experts, 8 active) and Flash at 309B / 15B active (48 layers, 39 sliding-window + 9 global attention), both natively omni-modal (text/image/video/audio), both 1M-token context, both MIT-licensed with weights on Hugging Face. Artificial Analysis scores Pro at 46 on its Intelligence Index — the highest open-weights result ever, ahead of GLM-5.3 (45) and Kimi K3 (44), up 20 points from V2.5-Pro’s 26, while still trailing closed leaders Fable 5.1 (53) and GPT-6 Astra (51). The gains came from scaled agentic RL executed as a 6-day public “Live RL” run: ~$2.62M (Pro) and ~$0.85M (Flash) of training compute, 30 steps each, ~750K trajectories, 1,568-sample batches with 1M-context training (2.7–3.7B tokens/step), MoE-router freezing and a multi-layer reward-hacking defense. Results: DeepSWE v1.1 Pro 58.4 → 72.57 (Flash 48.8 → 65.68), Terminal-Bench 2.1 89.9%, CyberGym 94.0%. The release includes 7,000+ open RL task environments (SWE, vuln reproduction, knowledge work, web design), the full verl/uni-agent/mini-swe-agent training stack, minimal composable harnesses, and a Distill-Qwen-9B that went 61.1 → 66.2 on SWE-bench Verified. API pricing is unchanged from V2.5 (Flash: $0.14/$0.28 per M on OpenRouter) — at matched intelligence, 1/20 to 1/60 of overseas closed models. MiMo Desktop shipped alongside, with ¥59–1,199 memberships and a 20× UltraSpeed mode. Notably, the model also closes manipulation loops in sim, controlling a Franka Panda arm from multi-camera vision.

Why it matters. This is the open-weights ceiling moving again — and the methodology being open-sourced matters more than the leaderboard position: task environments, reward design and harness-mixing recipes are the actual moat in the RL-scaling era, and Xiaomi just published its working set. Combined with yesterday’s Step 5 Preview ($1/M input) and Grok 4.7’s flat pricing, the pattern is unmistakable: RL compute is the new pretraining, it is getting cheaper per unit of capability, and the intelligence-cost Pareto frontier is now being pushed from three directions at once (closed labs, Chinese labs, open weights). Lei Jun’s ¥60B three-year AI pledge and the model’s Franka-arm demo hint the “brain” is being built for Xiaomi’s own physical-body roadmap.

3. Zhipu closes the ZCode data scandal: full open-source, third-party audits, “verify yourself” replaces “trust us”

What happened. The worst coding-tool trust crisis of the season reached its resolution arc (Sep 18–21). A developer discovered that Zhipu’s ZCode had been silently packaging and uploading entire workspaces — full Git history, LFS caches, configs — to an Alibaba Cloud OSS bucket whenever he was logged in, with no off switch; worse, a forensic review by enterprise user Taiyuan Chengming Technology found 6 workspaces uploaded (largest: 410MB, including what appear to be credential-adjacent material) even while privacy mode was on, and that uploads continued briefly after Zhipu’s first apology. Within three days ~400 developers organized; Zhipu’s stock (2513.HK) fell >5% intraday. The remediation, completed Sep 20–21, was unusually complete: the MaaS platform gained a “no data content retention” toggle; ZCode v3.14.0 removed the Repo Wiki feature and severed the local-repo snapshot/upload path entirely; the client, backend and Agent CLI were open-sourced to GitHub; and third-party audits by CAICT (China Academy of Information and Communications Technology) and NSFOCUS confirmed the zcode-prod OSS bucket and all data objects are deleted (“cloud zero data”) with no remaining exfiltration path. A standing vulnerability-report bounty was announced, code data was pledged never to have been used for training, and the stock recovered to close ~+2% on Sep 21. CITIC Securities’ read: open-sourcing converts unverifiable promises into verifiable ones — the product shifts from “believe the company” to “check it yourself.”

Why it matters. Coding agents sit on the most sensitive surface in enterprise software — repo histories, secrets, architecture. This is the first end-to-end incident response the industry can cite as a template: default-on “helpful” features + silent egress = the actual failure mode, and the fix that markets rewarded was transparency (open code + named auditors), not just deletion. Expect procurement checklists at every enterprise to now require data-flow disclosure, no-default telemetry, and auditability before any coding agent touches a monorepo. The unresolved tail — the Singapore-registered request endpoint vs the Beijing contracting entity, and the “deleted ≠ never leaked” problem — is exactly the cross-border data-governance question regulators will pick up next.

4. The IFR counts humanoids separately for the first time: ~7,000 sold in 2025 — the reality under the 890K forecasts

What happened. The International Federation of Robotics for the first time broke humanoid robots out as their own statistical category: ~7,000 units were sold worldwide in 2025 for industrial and professional-service use (consumer and military excluded; legs not required). The more revealing detail is where they went: IFR Secretary-General Susanne Bieller told Reuters many were bought by research institutions and companies generating data to improve AI models — not doing productive work — and even lead adopter automakers are running pilots with “single-digit or sometimes double-digit numbers of robots in their plants.” For scale: 542,000 conventional industrial robots were installed in 2024 alone, plus ~199,000 professional service robots. The full World Robotics 2025 dataset lands Sep 24. The census arrives amid a forecast arms race: Goldman Sachs raised 2030 global shipments from 256K to ~890K (2035: 6.5M; $138B TAM; 2026: 51K → 75K); Deutsche Bank lifted 2026 from 17.5K to ~50K (2030: 700K), calling China the core engine; BofA sees 90K in 2026 → 1.2M by 2030. On the ground in China, the scale-up is real but measured: Zhishen Technology (智身科技) passed 15,000 cumulative units produced (monthly capacity up 10× YoY, three bases/four factories, deployments in power, petrochemical, fire, security), H1 embodied-AI funding hit ¥93.5B (+5× YoY, 322 deals, >70% flowing to the top-20), and ~40 companies sit in the IPO pipeline (Unitree listed in August, Mech-Mind this month, Benmo Sep 29).

Why it matters. For the first time the industry has a hard, conservative baseline to measure hype against — and the gap is the story: 7K actual vs ~890K forecast for 2030 means every bull case depends on an adoption slope no robot category has ever achieved. The composition finding matters even more: the current market is data farms, not workers, which reframes today’s unit sales as training-infrastructure capex — precisely why “data rights” and dataset ownership (see item 6) are becoming the sector’s real balance-sheet items. Watch Sep 24’s full IFR release for the regional split; it will either confirm China’s majority share or reset it.

5. Coding agents move into the GUI — and Chinese open models colonize global clouds

What happened. Two distribution stories converged on Sep 21–22. Moonshot launched Kimi Code Desktop (macOS Apple Silicon/Intel + Windows, kimi.com/code): the Kimi Code CLI agent core wrapped in a graphical workspace with sessions, per-file/turn diffs, an agent-operable built-in browser with point-and-click annotation, embedded terminal, Git/PR status, and a four-tier control model — Plan (review-before-execute), Goal (long-horizon), Swarm (main agent dispatches subagents) and experimental Tower (parallel agents, explicit /tower opt-in) — plus three permission modes (Always Ask / Ask When Needed / Never Ask) and a secondary_model setting that delegates cheap subtasks to a smaller model. It runs on Kimi K3 (2.8T parameters, 1M context). Simultaneously, Amazon Bedrock added Kimi K3 (Sep 21) — the first 2.8T-parameter open model on Bedrock, with native vision, 1M context, and explicit prompt caching (a first for open weights on Bedrock) — making Moonshot the latest Chinese lab to monetize via overseas cloud revenue share. The demand side tells the same story: per OpenRouter (Sep 14–20), global weekly token volume hit 129T, with Chinese models at 67.46T (+10.28% WoW) vs US models at 14.21T (−34.7%) — the 21st consecutive week on top, and DeepSeek V4.1-Flash leading (+219% WoW).

Why it matters. Two structural shifts in one item. First, the coding-agent interface battle: CLI won execution, but GUI is winning trust — permission tiers, visible diffs and browser verification are how agents get unblocked inside risk-averse orgs (Kimi’s Tower requiring explicit activation is a quietly correct product decision about parallel-edit blast radius). Second, the go-to-market inversion: Chinese open-weight labs no longer need their own global cloud footprint — Bedrock/OpenRouter distribution plus revenue share is the new export channel, and call volume (not parameter count) is emerging as the metric that actually tracks real adoption. US-engine share falling 34.7% in a single week is noise-prone but directionally consistent with the price floors in items 1–2.

6. Jinbin’s “Employee 001”: Northeast China’s first industrial embodied robot clocks in — with the auto industry’s first data rights certificates

What happened. At a Sep 21 launch event in Shenyang, auto-parts maker Jinbin (金杯股份) rolled its first embodied robot off the line and handed it an employee badge — “Employee 001” — the first industrial embodied robot in Northeast China, while simultaneously launching the Jinzi Bianliang data factory with JV partner AVariable Robotics (自变量机器人) (JV formed March 2026). The structure is the point: Jinbin has 19 components/service subsidiaries and the Shenqi Group’s vehicle-manufacturing scenes behind it, so the plan is to deploy robots directly into its own sorting, handling, QC and inspection lines — “our own factories are the best testbed,” in VP Yu Bo’s words, inverting the usual build-then-find-a-scene sequence. The data flywheel is formalized: real-line data from the “Quantum-1” robot is cleaned/annotated/aligned/desensitized into training sets, and the Beijing International Big Data Exchange issued Northeast China’s first embodied-data rights certificates — the auto industry’s first formally 确权’d (title-registered) embodied datasets. AVariable’s dual-arm + standard-gripper sorting cell already runs at 1,911 pieces/hour at >98% accuracy on a top customer’s real line, with deployment cost ~70% below comparable overseas solutions. Jinbin’s “1+4” strategy frames the company as body manufacturer + component developer + data annotator + vertical-LLM builder, targeting a “mother port” of auto-industry embodied data spanning stamping through final assembly.

Why it matters. This is the cleanest template yet for answering embodied AI’s cold-start problem (no scene, no data, no iteration): a manufacturer with captive real scenes converts its own floor into a data asset, and the data-rights certificate turns that dataset into a transferable asset class rather than exhaust. If third-party buyers start licensing Jinbin’s datasets, the circular-revenue critique hanging over data-collection centers (item: Sep 21 brief) gets a market-based rebuttal. It also marks the quiet industrial pivot: forty-year-old parts suppliers becoming embodied-equipment companies because they own the one thing labs don’t — real production variance.

7. Hygon’s “compute reaches the physical world”: 1000-series edge CPU launches today in Shenzhen

What happened. Hygon Information (海光信息) — China’s flagship domestic x86-compatible CPU/DCU vendor (market cap ~¥561B) — officially launched its 1000-series CPU at a Sep 22 event in Shenzhen themed “Let compute reach the physical world”: a lightweight C86-architecture processor for low-power embedded and edge workloads, targeting robots, machine vision, smart manufacturing, industrial control, oil & gas, transportation and medical equipment — edge terminals, industrial devices and IoT nodes rather than datacenters. The 1000 series completes Hygon’s ladder (7000 datacenter / 5000 mid-range / 3000 cost-optimized / 1000 edge), with the pitch centered on what real-world deployment actually demands: real-time response, low power, industrial-grade reliability and security. The demand math it’s chasing: China’s edge-computing market reached ~¥114.6B in 2025 (hardware 71%), and 2025 industrial-robot output hit 773K units (+28%) — the world’s largest robot market needing local, deterministic, safe compute. Hygon’s own financials give it runway: H1 2026 revenue ¥9.10B (+66.5%), net profit ¥1.80B (+49.7%), R&D at 29.2% of revenue.

Why it matters. The sovereign-compute story has been a datacenter story (Ascend roadmaps, 50K-card clusters); this is the moment it descends to the robot’s chassis. A domestic vendor defining edge silicon around functional safety and real-time control — not recycled server dies — closes another layer of China’s “silicon + model + data infrastructure + factory deployment” full stack for embodied AI, and pressures robot OEMs to declare compute roadmaps the way EV makers once declared battery ones. The honest caveat: embedded edge is a crowded field (Phytium, Rockchip, Allwinner, NXP), and customer qualification cycles for industrial silicon run in years, not keynote cycles — watch for anchor customers and mass-production dates, not the announcement.

8. The “enter-key engineer” reckoning: a 5M-view post, and AI now writes 1-in-6 Linux kernel patches

What happened. The human-cost debate of agentic coding went mainstream. A post by engineer “v0xium” (Anonymous, via Business Insider, Sep 21) — describing a new job where Claude Code produces the specs, tests, PRDs, tickets and reports, engineers work “12–13 hours a day just pressing enter,” nobody reviews deeply, and management measures sprint velocity — hit ~5 million views, drawing Elon Musk (“Yikes”), Chamath Palihapitiya (warning of a generation turned into “people pressing buttons at casino slot machines”) and GitLab CEO Bill Staples (“Humans have become the orchestration layer of machines”). The measurable companion data from the same window: AI-generated code accounted for 17.25% of Linux kernel patches in September (~1 in 6, per Lunduke Journal’s figures); Linear reported AI coding surged commit volume until CI became the bottleneck and had to re-engineer its pipeline; one viral trace showed Opus 5 spawning 17 subagents and 1.5M cache reads to answer a trivial model-lookup; and Berkeley’s “Harness Tax” work found lightweight harnesses often beat heavyweight scaffolds on both score and cost. Artificial Analysis’s Coding Agent Index now shows Claude Code, Devin and Codex all at 62 — prompting the week’s sharpest observation: the question is no longer “can AI write code” but “how much of the development loop can it close.”

Why it matters. The first-order bottleneck has decisively moved: not generation (solved, cheap, and increasingly accountable for a sixth of kernel patches) but review capacity, CI throughput and organizational process — exactly what the Real-SWE benchmark’s 38.8%-on-private-code finding predicted last week. The v0xium post is the cultural half of that data: velocity-metric-driven adoption without review slack converts engineers into unverified-diff forwarders, and the backlash is arriving before the productivity numbers are settled. For teams, the actionable takeaway from the same threads is small but real: ~40 lines of directive CLAUDE.md and lighter harnesses consistently outperform heavy scaffolding — governance and context hygiene are now performance features, not compliance overhead.

Also on our radar

• Agibot × Chimelong embodied theme park opens Sep 24 (Hengqin): 300+ Agibot robots across the park on day one — the first mass deployment of embodied AI in a high-footfall public entertainment venue, and a live stress test for multi-robot coordination and maintenance economics.
• Humanoids Summit Seoul, Sep 22–23 (COEX): Korea’s MSIT participates with a “Physical AI Policy in Korea” talk; NVIDIA × ROBOTIS workshop; speakers from Persona AI, FieldAI, RLWRLD, Unitree APAC — Korea formally entering the physical-AI state-strategy club.
• Unitree Dex5-S dexterous hand (Sep 21): 22-DoF, 1:1 human-hand size, ~620g, from ¥39,900; Pro adds tactile sensing — the component-level price curve keeps falling.
• TypeSafe’s Jev goes GA (no waitlist; ~120M free tokens; $0.042/M input, output free): the typed decision-model layer (Choice/Score/Now, 70–500ms) that Vercel uses for 5–18× classifier speedups — the “small-decision model + big-generation model” split is becoming a default agent architecture.
• HarmonyOS coding stack: Alibaba’s Qoder became the first AI-coding agent workbench on HarmonyOS PC (Sep 21), while Huawei Cloud CodeArts shipped a HarmonyOS-specific coding model (11M+ developers; error rate −80%, first-pass compile +78%) — vertical-ecosystem developer tooling as strategy.
• FutureX papers of the day: PARTS — real-robot RL restricted to bottleneck subtasks lifts Franka long-horizon success 50% → 95% (dual-arm YAM 32% → 61%) at tens of minutes of rollouts per task; VLA-Feedback — repurposing the last diffusion-denoising step as a high-frequency corrective loop lifts dynamic-task success 27.5% → 85% in sim, 51% → 73% on real robots.
• Vals AI’s 141-hour Minecraft trial of GPT-6 Astra: flawless long-horizon planning until catastrophe — after its base was creeper-bombed, the agent regressed into a perpetual potato-farming loop instead of replanning. “Frustration recovery” is emerging as the next benchmark frontier for long-running agents.
• fastbrowse (MIT license): deterministic browser-agent harness scoring 41/42 on official evals at $0.0057/task vs $0.4070 for hosted Browser Use — ~71× cheaper, the cost-engineering theme (item 1) applied to web agents.

Closing Take

The day’s releases all make the same move: hold the price, raise the capability — Grok 4.7 at $2/$6, MiMo-V2.6 at 1/20–1/60 of closed rivals, Jev effectively free per decision. That deflation is colliding with two honest countweights: the IFR’s 7,000-robot census telling the embodied world its forecasts rest on a 100× leap of faith, and a 5M-view post telling the coding world its productivity story has an unbilled labor line called review. The teams compounding fastest are the ones building the unglamorous middle layers — data rights, audit trails, permission tiers, CI capacity — while everyone else argues about leaderboards.

Curated & analyzed by @WoLoveAI · Sources triangulated across xAI/Artificial Analysis/Cursor technical disclosures, Xiaomi MiMo official release, The Paper/Securities Times/Toutiao/Phoenix Tech, Reuters/IFR, Economic Information Daily, Shenyang Daily, FutureX Physical AI Daily #127, AGI Hunt, AlphaSignal, Business Insider, and IT Home.