EAIDaily — 2026-09-21

English AI Daily Brief | Focus: AI Coding & Embodied Intelligence Automation series, Issue #16 · Coverage window: Sep 20–21, 2026

Today’s Master Frame

The bill for the “slow down” consensus arrives — and the checkout counter opens for robots. On the coding side, last week’s frontier-lab coordination on pacing AI now has a courtroom test: paying users filed an antitrust class action, while market data (OpenRouter spend crossover, Ramp enterprise-spend shares) shows the competitive clock that makes “slowing down” expensive. Token prices hit new floors (StepFun: $1/M input). On the embodied side, robots crossed from data centers into living rooms and theme parks — a ¥19,999 consumer humanoid with day-one shipping, 300+ Agibot robots entering a theme park on Sep 24 — and the data economy behind them got its first public price list (¥500–1,000/hour for real-robot data). Everything today is about pricing: of tokens, of data hours, of robots, and potentially of the slowdown itself.

1. Antitrust class action targets the “AI slowdown pact” — safety coordination enters the courtroom

What happened. Four paying users of ChatGPT, Claude, Grok and Gemini have filed a proposed class action in the U.S. District Court for the Northern District of California (filed Sep 18, AP-reported Sep 19, dominating the Sep 20–21 news cycle) accusing OpenAI, Anthropic, Google and SpaceXAI/xAI of violating the Sherman Antitrust Act by coordinating to decelerate frontier AI development. The complaint anchors on Dario Amodei’s September 12 “pace the frontier” essay and the same-day public endorsements from Sam Altman, Elon Musk and Demis Hassabis; it also cites a July statement by AI-company executives calling for a government-backed global slowdown (admitting “huge competitive pressure” made unilateral slowdown unlikely), and reports that OpenAI executives have acknowledged weeks of contact with Anthropic and Google on AI safety. Lead counsel Nick Rowley argues the pact degrades the value of paid subscriptions; the companies had not responded as of Sep 19.

Why it matters. Safety rhetoric is now a legal discovery surface. Whatever the merits, the suit establishes a new governance mechanism nobody planned for: antitrust litigation as the de-facto enforcement layer over frontier-lab coordination — arriving precisely as Washington declines to legislate and labs prepare mega-IPOs (Anthropic ~$2T, OpenAI ~$1.2T private round). If “everyone agreed to slow down together” is provable, it re-prices the entire safety-communications playbook; if not, the discovery process itself will map how labs actually talk to each other.

2. Anthropic’s counterpunch: a new model in safety evaluation, and the money behind the slowdown talk

What happened. Per multiple reports (Sep 20–21), Anthropic is evaluating a new model to counter OpenAI’s Astra, still inside safety assessment while debating how to lift monetization alongside launch — six days after Amodei publicly called for the industry to slow capability gains. The competitive data explains the urgency: Ramp tracks OpenAI’s new model at ~13% of enterprise AI spend vs Anthropic’s ~8%; OpenRouter says user spend on OpenAI models exceeded Anthropic’s for the first time last week; annualized run-rates stood at >$65B (Anthropic) vs >$40B (OpenAI) as of end-July. Meanwhile the capex machine accelerates: Anthropic quietly built a physical “wet lab” in the Bay Area (Claude commanding lab robots, confirmed by its life-sciences lead) and signed a ~2.16 GW data-center lease in Australia (Zerra DC, online from 2027, mainly inference). OpenAI separately projects up to $278B in cumulative negative free cash flow through 2030 against a $350B revenue target.

Why it matters. The contradiction is now measurable: labs that call for pacing are signing gigawatt leases no governance timeline can pause — power and grid schedules don’t wait for safety evaluations. The OpenRouter crossover is the first demand-side signal that “frontier ≠ default choice,” and it pressures Anthropic to ship under evaluation rather than slip dates. For the coding audience: the model that wins enterprise spend share becomes the default agent brain in production systems — worth watching more than any single benchmark.

3. StepFun Step 5 Preview: a 600B agentic MoE at $1 per million tokens — open weights promised in 25 days

What happened. On Sep 20, Shanghai’s StepFun launched Step 5 Preview, its new flagship for long-horizon agentic work (coding, software engineering, finance): a sparse MoE with 600B total / 27B active parameters (~4.5%), a 92-layer narrow-deep stack, 1M-token context, text+image input, and aggressive pricing — $1.00 per million input / $2.70 per million output tokens, with 95% cached-input discount (~$0.05/M), ~100 tok/s. Technical claims: on-policy long-horizon RL with bit-wise train–inference alignment, MTP-3 speculative decoding, FP8 MoE, KV-cache offload (>3× end-to-end RL speedup); Sparse Grouped-Query Attention with block-wise token merging cuts indexer/top-k cost to ~1/8. Artificial Analysis scores it 44 — tied with Kimi K3 (max) at ~65% lower cost per task; Terminal-Bench 4.0 33% (vs DeepSeek V4.1-Flash’s 27%), SciCode 59%. API is live now; full open weights scheduled for October 15 (license unspecified). Early community evals: strong scientific coding, gaps in factual recall and complex-document handling. xAI also put Grok 4.6 on its API the same window (500K context, selectable reasoning effort, $2/$0.50/$6 per M tokens).

Why it matters. The agentic “Pareto frontier” is now explicitly a pricing battleground: capability-per-dollar, not raw scores, is the launch thesis. A 1M-context agent model at $1/M input makes long-horizon coding sessions (repo-scale context, tool-loop transcripts) financially trivial — and an open-weights commitment 25 days out keeps pressure on GLM/Qwen/DeepSeek. The same week Anthropic considers raising prices to fund IPO-era economics, the commodity end of the market keeps collapsing. Enterprise buyers get a real arbitrage: frontier for hard tasks, this for volume agent loops.

4. Agibot’s consumer beachhead: Qiyuan Q1/T1 personal robots launch at ¥19,999 — “on sale means shipping”

What happened. On Sep 20 in Shanghai, Qiyuan (启元) — the consumer embodied-AI brand of A-share listed Sunwo New Materials (上纬新材), which Agibot (智元) acquired in a reverse merger; Agibot co-founder/CTO Peng Zhihui (“Zhihuijun”) is Sunwo’s chairman — launched its first two personal robots with same-day sales start and shipping from Oct 1: Q1 and T1 both at ¥19,999 (Q1 Explorer ¥26,999; T1 Pro ¥29,999), with experience stores across 7 cities. Q1 (88 cm, toddler-height) is positioned as kid companion / tech collectible / developer platform — 3D-printable shells via the PrimeFLOW design pipeline, customizable personality/actions/voice, open SDK/HDK on the Explorer tier; its QDD quasi-direct-drive joints are egg-sized (~260 g). T1 is billed as the world’s first transformable personal robot — autonomously switching between wheeled-bipedal and quadruped forms; T1 Pro adds NVIDIA Orin Nano, 360° obstacle avoidance, auto-recharge and voice summoning. Software stack: PrimeMotion (safe compliant motion), PrimeGo (multimodal following), PrimeVista (interaction), with a household foundation model PrimeMind and a Harness-architecture robot agent on the roadmap; Tencent’s WorkBuddy agent is integrated via A2A invocation / connector binding / plugin composition behind a four-layer physical-safety chain (permission control, action whitelisting, simulation preview, on-site confirmation). H1 2026 R&D spend was ¥180M (¥160M on consumer embodied; 22.5% of revenue), with ¥500M planned for the full year. Chief scientist Dong Hao (PKU): embodied scaling laws exist but differ from LLMs — the strategy is to push 3 tasks to ~100% first, then expand, entering through companionship rather than chores.

Why it matters. The consumer price floor just dropped an order of magnitude: ¥19,999 (~$2,800) vs UBTECH U1’s ¥119,800 and Unitree G1+ at ¥95K — and unlike demo-stage rivals, this is buyable-and-shippable hardware with an explicit developer-open tier, which is how a data flywheel actually gets seeded (Q1 Explorer + open SDK = user-generated skills and telemetry). The reverse-merger structure is equally notable: a listed company as the consumer-robot vehicle gives Agibot a capital-markets arm while the parent stays private. “Companion first, chores later” is the same sequencing thesis Figure’s critics are about to test (item 7).

5. Agibot × Chimelong: 300+ robots enter a theme park — the largest public-scene stress test yet

What happened. On Sep 24, the world’s first large-scale embodied-intelligence theme park, built by Agibot (智元) and Chimelong Group, opens at Chimelong Spaceship Park in Hengqin (Zhuhai): over 300 robots across Agibot’s full product line go on duty simultaneously on opening day.

Why it matters. This is a different kind of deployment than the factory pilots dominating headlines: a high-traffic, high-interaction public venue is a mass stress test for multi-robot coordination, human-robot interaction, fleet ops and maintenance economics — plus a continuous real-world data loop. Coming one week after UBTECH’s first U1 consumer deliveries and four days after Qiyuan’s launch, it completes a pattern: Chinese embodied-AI players are attacking the B2C gap from both ends (homes + entertainment venues) rather than waiting for industrial ROI alone. The open question is whether it becomes a sustainable operating IP or a very expensive demo — cost, reliability and experience differentiation will decide, and Agibot now has 300 units’ worth of evidence incoming.

6. China’s embodied-AI data economy gets a price list: 106 collection centers, real-robot data at ¥500–1,000/hour

What happened. A compiled industry census (Huxiu, Sep 21) counts 106 centralized embodied-AI data collection centers in China as of August 2026, 84 already operational, housing ~3,972 disclosed robots (likely >5,000 running). Scale is still small: 55.8% of centers deploy ≤50 units; the two largest (Yibin High-Speed-Rail Southwest Training Center, Kaipuyun Training Ground) run 300 each; Jiangsu leads with 18 centers, then Zhejiang 17, Guangdong 11. Construction accelerated from 2.7 new projects/month (Mar 2024–Mar 2025) to 7.3/month (Apr 2025–Jan 2026), peaking in Q1 2026. Teleoperation is the dominant collection method (79% of centers). For the first time, pricing bands have formed: real-robot data ¥500–1,000/hour, UMI handheld data ¥300–400/hour, simulated data ¥200–500/hour. Platform concentration: Agibot appears in 31 centers, LimX Dynamics 15, UBTECH 13. The census lands amid the Mech-Mind CEO’s public accusation that collection centers mask circular revenue via related-party transactions (Galbot filed a police report denying it).

Why it matters. Data supply chains now have quotes — that’s how you know a market, not a subsidy program, is forming. Real-robot data commanding a 2–5× premium over synthetic is the market’s verdict on where the scaling-law bottleneck sits (echoing Figure’s Helix claim and its critics). But the circular-revenue dispute is the caveat that matters: if a large share of “demand” for data hours comes from the platform vendors themselves, today’s ¥500–1,000/hour is a transfer price, not a market price. Watch whether third-party buyers (auto, consumer-electronics, household-appliance OEMs) start appearing in these centers’ order books.

7. The embodied scaling-law debate turns adversarial — and GPT-6 Astra drives real arms directly

What happened. Two items from the weekend’s physical-AI discourse:

• Peers push back on Figure’s Helix 2.5 claim. Following Figure’s “first measured human-to-robot transfer scaling law” (56% combined success across 30 unseen homes vs ~9% baseline), Sunday Robotics co-founder Tony Zhao noted that 237 successes out of 420 trials “means the robot is failing nearly half the time”; stealth-founder Nikolai Ensslen was blunter — the data “proves that the imitation-learning system Figure uses cannot generalize.” A key caveat on all sides: the training data for all three tasks was still prepared by Figure itself — what Helix 2.5 never saw was only the test residences, layouts and objects.
• GPT-6 Astra as a direct visuomotor policy (single-source, unverified). The RoboCurve evaluation reportedly connected GPT-6 Astra to a pair of real I2RT YAM robotic arms: the model views three camera feeds (overhead + two wrists) plus arm state, and directly outputs end-effector poses (x, y, z, yaw, pitch, roll, gripper), with inverse kinematics handling joints — scoring 19/20 on red-block pick-and-place.

Why it matters. Embodied AI is acquiring the adversarial peer-review culture that discipline in AI coding benchmarks took years to build: the same 56% reads as “scaling law confirmed” or “fails half the time” depending on your benchmark philosophy — and “unseen homes” ≠ “unseen tasks,” which keeps Figure’s data-moat story intact but narrows it. The Astra-on-arms claim, if it survives scrutiny, is the bigger earthquake: a frontier LLM bypassing dedicated VLA training entirely and acting as its own low-level policy would change the data-moat math for every collection center in item 6. Treat it as a signal to verify, not a result to cite.

8. Hygon to unveil a “new-category” chip for the physical world (Sep 22, Shenzhen)

What happened. Hygon Information (海光信息) — one of China’s two leading domestic CPU/DCU vendors — will release a new-category chip for physical-world scenarios on Sep 22 in Shenzhen, targeting robots, machine vision and intelligent manufacturing applications. No specs disclosed pre-launch.

Why it matters. The compute layer for embodied intelligence is getting its own dedicated silicon strategy from a flagship domestic vendor — not a repurposed GPU, but a category defined by real-time, low-power, safety-critical edge workloads (the same requirements Toyota’s ELEY and Figure’s onboard stacks wrestle with). If credible, it closes another gap in China’s “model + data infrastructure + silicon + factory deployment” full stack, and it will force the robotics OEMs to declare compute roadmaps. Watch for: real compute/watt figures, anchor customers, mass-production timing, and whether a software ecosystem (toolchain, ROS integration) ships with it.

Also on our radar

• Trump announces an “AI Force” (Sep 19): a federal AI task organization plus a soon-to-be-named AI “czar,” with an AI-could-be-25%-of-GDP framing and a pledge not to hinder development — concrete mandate and budget unspecified.
• Huawei Cloud CodeArts Agent goes commercial in Malaysia (Sep 18 launch, reported Sep 20): enterprise coding agent with a 16-specialized-agent team (requirements→architecture→coding→testing→review), spec-driven development, 30+ reusable Huawei engineering skills, and comprehension of tens-of-millions-of-line codebases — Chinese agent tooling scaling into ASEAN as the region’s data-center capacity (1.5+ GW in Johor/Klang Valley) comes online.
• Agility Robotics sets Analyst & Investor Day for Oct 6 ahead of its Q4 SPAC close (S-4 filed Sep 9; Nasdaq ticker “AGLT”): first pure-play humanoid on U.S. public markets, with Digit 5 (50 lb payload, 90-min battery, 9-min recharge), 65,000+ logged hours, and ~$300M in multiyear orders — the mirror-image listing to Unitree’s STAR Market debut.

Closing Take

One week after the industry’s CEOs agreed in public to slow down, the week’s data says everything is speeding up: a lawsuit over the slowdown itself, a spend crossover forcing Anthropic’s hand, token prices at new lows, and robots shipping to living rooms, theme parks — with their data now carrying hourly quotes. Governance is being built out of courtrooms and price discovery, not white papers.

Curated & analyzed by @WoLoveAI · Sources triangulated across AP/WSJ-derived coverage, 每日经济新闻, 华尔街见闻, 券研社, 上海证券报, 科创板日报, 64bit Daily Brief, FutureX Physical AI Daily #126, Huxiu, Pandaily, Artificial Analysis, and vendor announcements.