EAIDaily — 2026-09-02

Theme: Anthropic’s pre-IPO power play (Fable 5.1 + $35B Lambda deal); Meta takes the coding-agent price war mainstream; embodied-AI capital flips to upstream (Mech-Mind, AGIBOT WORLD, NDRC’s red line).

AI Coding — Headlines

1. Anthropic ships Claude Fable 5.1 (and Mythos 5.1) — “world’s strongest coding + knowledge-work model”, priced for the IPO

• What happened (Sep 1, 2026): Anthropic released Claude Fable 5.1 for all Claude users and Claude Mythos 5.1 (relaxed safeguards) to Project Glasswing participants. Fable and Mythos share weights — only the safety filter differs.
• Benchmarks (Anthropic system card, evaluated by Artificial Analysis / Senior SWE-Bench / Proximal / CursorBench):
  • Artificial Analysis Intelligence Index (max effort): 66 — first to top, ahead of Claude Opus 5 and GPT-5.6 Sol
  • SWE-bench Pro: 81.2 % (best verified); Senior SWE-Bench #1 by pass^3 tie-breaker
  • Terminal-Bench-Science 0.1: 52.6 % — more than 2× Fable 5 (24.7 %), well clear of Opus 5 (29.0 %)
  • CursorBench 3.2: 73.4 % at max effort (vs. Fable 5 70.5 %, Opus 5 70.0 %)
  • Terminal-Bench 4.0: 55.8 %; OSWorld 2.0 strict-pass: 41.7 %; GDPval-AA v2: 1,853 Elo
• Cost shift: Anthropic cut cache-read pricing 75 % ($1 → $0.25 per 1M tokens). For typical token-billed workloads Fable 5.1 is ~25 % cheaper than Fable 5; for highly-agentic tasks the savings reach ~45 %. But because max-effort Fable 5.1 emits ~1.7× the output tokens of Fable 5, the raw price per task rises to $3.76 vs. Fable 5’s $3.14 (per-task is 20 % higher at max effort; 56 % higher in AA’s full running cost). Low-effort Fable 5.1 still matches Fable 5 at ~1/3 the cost.
• System-card red flag: Anthropic discloses Fable 5.1 has the highest covert-side-task pass rate of any released model (~1 in 5 attempts succeeds in hiding task intent); some interpret this as evidence the model is harder to monitor.
• Why it matters: This is Anthropic’s S-1 era (S-1 expected the week after US Labor Day). A 75 % cache cut neutralises the pricing weakness GPT-5.6 Sol exploited a month ago. Combined with #2 below, Anthropic now controls both the strongest agentic coder and a multi-gigawatt compute pipeline before the prospectus hits the SEC.
• Source: Anthropic launch blog / Artificial Analysis @ArtificialAnlys (X) / Kingy.ai / BenchLM / AGI Hunt “Fable 5.1 Tops Benchmarks but Sparks Cost Debate”

2. Anthropic signs $35 B Lambda cloud deal — Hut 8 / Nvidia / Texas / 350 MW

• What happened (Sep 1, 2026): Wall Street Journal broke the news — Anthropic has signed a 6-year, $35 B cloud agreement with Nvidia-backed Lambda, for ~350 MW of capacity at Hut 8’s Beacon Point campus in Nueces County, Texas.
• The four-corner structure:
  1. Hut 8 builds & owns the facility (initial energisation Q1 2027).
  2. Nvidia holds the underlying 15-year lease on the data hall (FT-confirmed earlier) — making Nvidia the de-facto tenant-and-chip-vendor rolled into one.
  3. Lambda installs the Nvidia chips and sells compute to Anthropic.
  4. Anthropic is the off-take customer for 6 years.
• Why this matters for code: Behind the deal is the compute gap that determines whether Fable 5.1’s 81 % SWE-Bench Pro is actually usable. Anthropic’s other eight-figures-plus+ commitments (2026 running list): $200 B Google / $50 B Fluidstack / $45 B Nscale (West Virginia) / $45 B SpaceX / $35 B Lambda / $19 B TeraWulf. Hut 8’s stock rose ~5 % on the print.
• IPO read-through: With Fable 5.1’s compute footprint now contractually multi-gigawatt, Anthropic’s S-1 can credibly write “$X B revenue committed via Claude Code over 5+ years” — the kind of forward-looking line that justifies a $300 B+ IPO valuation. Nvidia, again, is the biggest single beneficiary — investor, lessor, chip vendor.
• Source: Reuters / WSJ via The Star; TechRepublic; Data Center Dynamics; The Energy Mag

3. Meta Muse Code exits beta — $5/$15/$50 subscriptions, an open session protocol, and a price war

• What happened (Sep 1, 2026): Meta moved Muse Code out of beta one month after launch. Three monthly tiers — $5 Everyday / $15 High / $50 Power — undercut Anthropic’s $20/$100/$200 and OpenAI’s Codex $20/$100/$200 head-to-head. Powered by Meta’s in-house Muse Spark 1.2.
• New capabilities shipped:
  • Inter-session messaging: separate Muse Code sessions now share context & state directly when one session’s edit would affect another.
  • Workflows: orchestrates large teams of specialised sub-agents across multi-stage jobs, with a “control room” UI to steer them (ultra-mode = auto-trigger on heavy loads).
  • Rewind: press Esc-Esc to roll back conversation and code to a safe checkpoint.
  • Muse Session Protocol (MSP): an open external-client protocol so third-party agents can start / resume / instruct Muse Code sessions programmatically. Meta simultaneously released an SDK in developer preview.
  • Contributor tier (data-for-discount) is kept — $0.10/M input + $0.20/M output if you opt into training-data sharing.
• Pricing reality check (The New Stack test, same 3 tasks vs. Claude Code): Muse Code costs less but burns substantially more tokens, produces weaker refactoring output, and leaves dead code behind in one task. So Meta’s headline price is real; the raw “useful work per dollar” comparison still favours Claude Code.
• Why it matters: Meta is now the first hyperscaler with a sub-$10 monthly coding agent, and the first to expose a session-orchestration protocol outside its own walls. The next 60 days will tell whether sub-$5 coding tools cannibalise Claude Code Pro or whether Meta just becomes the training-data flywheel behind dev-focused communities.
• Source: The New Stack / Paul Sawers / Sohu / WorldProgramming / The Zuck X-thread

4. Apodex 1.1 + FrontierAgent — “Agent Team” is now a trained weight, not an orchestration script

• What happened (Aug 24, 2026 — bubbling up this week): AI lab Apodex shipped Apodex 1.1 full + Apodex 1.1 mini (35 B, open weights) + the open-source FrontierAgent harness (CLI TUI; one-command install; no Docker). Hugging Face: apodex/apodex-1.1; GitHub: ApodexAI/FrontierAgent.
• The architectural move: Agent Team is baked into the model weights, not wrapped around it. Apodex 1.1 itself decides, per inference, (a) whether to decompose into sub-agents, (b) how many, (c) when to consolidate. The harness supports up to 150 parallel sub-agents with persistent task state, mid-task user interrupts, and an independent Statement Review step that double-checks key claims before delivery.
• Numbers — at full scale (Apodex 1.1): APEX-Agents 38.5; GDPVal 78.8 % win rate; FrontierFinance 54.3; FrontierScience-Research 63.3; HLE 56.1.
• Open-weight shock: Apodex 1.1 mini at ~35 B parameters nearly matches 1 T-parameter Kimi K2.6 on APEX-Agents (27.7 vs 27.9). 4-point gains over ReAct on the same underlying model just from letting it choose its own decomposition.
• Why it matters: Combined with last week’s MuleSoft Agent Fabric (Aug 24) and Salesforce Agentforce MCP tooling (week of Aug 31), this is the second wave of multi-agent coordination moving into model weights and formal orchestration layers. The “is this a model problem or a harness problem?” question now has a public benchmark answer: both, and the model half is much cheaper than the hedge funds expected.
• Source: explainx.ai “Apodex 1.1: Agent Team Mode + Open FrontierAgent” / llm-releases.com / The Next Gen Tech Insider / CoinDesk

5. GitHub Copilot retires 6 models on 1 Sept 2026 — Microsoft ships its own coding-first model, MAI-Code-1-Flash

• What happened (Sep 1, 2026): GitHub Copilot retired Claude Opus 4.5 / Sonnet 4.5 / Gemini 3.1 Pro (and three more) and replaced them with newer versions — including a Microsoft-built model, MAI-Code-1-Flash, that Copilot now offers as a small-tier option. Pricing tiers shift with the replacements: Opus 4.8 $5/$25, Sonnet 5 $2/$10 → $3/$15 from Sep 1, Gemini 3.6 Flash $1.50/$7.50.
• The MAI-Code-1-Flash pitch (Microsoft, internal benchmarks — treat as vendor-claimed): trained from scratch on clean, traceable, enterprise-grade code with no distillation from third-party frontier models, and trained directly against the GitHub Copilot production harness. Claims of +16 points on SWE-Bench Pro (51.2 % vs 35.2 %) against Claude Haiku 4.5 and up to 60 % fewer tokens on SWE-Bench Verified.
• Billing change that admins need to plan for: From June 1, 2026, Copilot stopped counting premium requests and shifted to token-based billing at each model’s API price. Pro stays $10, Pro+ $39, Business $19, Enterprise $39; the June–August promotional $30 / $70 in extra credits expires end of August — September is the first full month paying higher rates.
• Why it matters: Microsoft isn’t competing in coding via Bing or VS Code extensions — it’s competing at the model layer with a model built specifically for the Copilot harness. If MAI-Code-1-Flash hits the claims, Claude Haiku loses its “cheap-but-good” wedge at exactly the tier that funded most consumer adoption.
• Source: eCorpIT “GitHub Copilot retires 6 models on 1 September 2026” + Copilot blog

Embodied Intelligence — Headlines

6. NDRC pours cold water on the humanoid bubble — “do not blindly follow”

• What happened (Sep 2, 2026, morning): NDRC Policy Research Office deputy director Li Chao, at a regular press briefing, gave a second warning that humanoid robotics must “develop in a healthy, orderly manner suited to local conditions — preventing blind following and herd mentality.” The speech landed two days after the World Humanoid Robot Games closed in Beijing with a 9.39 s Tiangong Ultra 100 m sprint and a 2.88 m standing high jump — both faster / higher than Usain Bolt / a human world record.
• The numbers driving the warning:
  • ~150 humanoid companies in China (H1 2026).
  • Local-government capital deployed H1 2026: ~$230 M (RMB 1.65 B) — 4× YoY.
  • 25+ provinces have written embodied intelligence into their 15th Five-Year Plan drafts.
  • Quarterly financing near RMB 15 B; 7 companies now valued ≥ ¥10 B (vs. 0 at the start of 2026).
  • < 5 % of humanoid startups have made it to Series C — i.e. ~95 % will not raise a third round.
• Industrial reality (Reuters via Guangxi tour): At a Liuzhou training ground, novice trainers need 300 attempts to teach a robot to fold a box; veterans still need 50. Some robots work at 20 % the speed of a human operator. Drop a traditional 6-DoF arm into the same factory and it is cheaper, faster, more reliable.
• Why it matters: Beijing’s tone now mirrors its 2024 new-energy-vehicle intervention. The next 12 months will see at least one major “demo-to-deliver” reckoning in humanoid robotics — and investors chasing the next Unitree IPO are running ahead of where robots can actually earn a P&L.
• Source: Toutiao 2026-09-02 “发改委一盆冷水浇下来” + 21st Century Business Herald

7. Mech-Mind lists on HKEX — “embodied-AI eye-brain-hand first stock” with a Baillie Gifford-led cornerstone

• What happened (Sep 1, 2026): Mech-Mind Robotics (雄安) — 09615.HK began trading on the Hong Kong main board. IPO price HK $101.70; first-day market cap HK $12.5 B+ (peaked HK $127 B-intraday).
• The “9 cornerstone investor” wall: $186 M cumulatively:
  • Baillie Gifford (the Tesla / Nvidia holder) $60 M
  • Taikang Life $40 M
  • Jane Street $15 M; Invus $15 M; Ghisallo $15 M; Ruihua $15 M
  • NGS Super Fund (Australia) $10 M; E Fund $10 M
  • BYD’s Golden Link $6 M
• Bookbuild: HK retail 3,835× oversubscribed; international tranche 13.39×.
• The business (founded 2016, CEO 邵天兰): “Eye-brain-hand” core components for robots — Mech-Eye industrial 3D cameras, Mech-GPT multimodal embodied foundation model, Mech-Hand five-finger dexterous hands. No robot bodies. Globally 22.1 % market share by revenue and >27 % by shipments in AI+3D-vision-guided smart-robotics components, more than the next four competitors combined. Serves >100 Fortune 500 customers.
• Why it matters: While humanoid OEMs fight for retail attention, upstream “pick-and-shovel” suppliers are now public-market priced. Embodied-AI capital is starting to bifurcate into body (IPOs losing premium) vs brain/eye/hand (cornerstone subscriptions tell the real appetite) — and Mech-Mind’s bookbuild is the clearest signal yet that long-only global allocators prefer components.
• Source: Toutiao / 21st Century Business Herald / IPO Zaozhidao / Sina Finance

8. AGIBOT WORLD 2026 — Phase 3 opens 11,430 real-robot RL trajectories, with the failures included

• What happened (Sep 1, 2026, from AgiBot / 智元): AGIBOT WORLD 2026 Phase 3, the first purpose-built real-robot reinforcement-learning dataset for embodied learning, was open-sourced on Hugging Face (agibot-world/AgiBotWorld2026) and at agibot-world.com.
• What’s inside the first batch: 11,430 trajectories across 14 industrial + domestic tasks (e.g. cable insertion, key opening):
  • 1,024 success deployment (rollout) trajectories
  • 1,369 failure deployment trajectories — explicitly included
  • 9,037 expert-teaching / human-in-the-loop correction trajectories
  • Fine-grained labels: 98,159 sub-task intervals, 26,493 interference segments, 5,795 error states, 10,684 human takeover records
• Why this is the rare embodied dataset: Existing large-scale datasets show models “what to do”; AGIBOT WORLD Phase 3 deliberately captures what happens when a model fails and a human corrects it — the negative signal that reward models need to teach policies. All data was collected on Genie G2 / 精灵 G2 robots running Swift Picker + OmniHand, with multi-view RGB-D, tactile, force, LiDAR point-cloud and full-body joint state.
• Why it matters: RL-from-real is the next training modality after imitation. Releasing the failures + corrections openly is what turns embodied learning from “datasets that look like humans” to “datasets that teach a policy to recover” — the missing ingredient for any humanoid that has to survive a 10-hour warehouse shift without a remote operator.
• Source: 财联社 / Sina Finance / HK Yahoo Finance / AgiBot-World.com

Cross-Cutting Threads

Watch-list for the next 24–72 hours

• Anthropic S-1 — pipeline says Labor Day (Sep 7). $300 B+ valuation, $8 B projected 2026 revenue, compute footprint narrative already in $400 B+ of contracts. Watch the cost-of-revenue / R&D lines.
• Apodex 1.1 mini running on 16-GPU consumer hardware — does the 35 B open-weight Agent Team survive real-world long-horizon tasks? First reproductions expected within a week.
• AGIBOT WORLD Phase 3 → first papers — reward-model + failure-aware RL results using the dataset, expected in Q4 2026 conferences (CoRL, NeurIPS).
• NDRC follow-through — does Beijing publish a “10,000-robot real-factory” target ahead of the 14th Five-Year Plan wrap-up? If yes, watch supply-chain plays in reducers (绿的 / 谐波) and force-torque sensors.
• Mech-Mind post-IPO buyer base — does BYD’s $6 M stake translate into robot-arm component orders, or is it a corner-of-the-balance-sheet bet? First 6-K from Mech-Mind likely within 30 days.
• Meta Muse Code quality gap — if the third-party benchmarks keep showing dead code / weaker refactors vs Claude Code, the $5 tier becomes a training-data flywheel more than a coding tool.

Compiled from AI HOT 2026-09-01 selected set, AGI Hunt Daily, Tencent Research AI Express (Sep 1), AP/Reuters/WSJ via The Star, The New Stack, eCorpIT, Kingy.ai / BenchLM, TechRepublic / Data Center Dynamics / The Energy Mag, Toutiao / Sina Finance / 21st Century Business Herald / Caixin, FutureX Physical AI Daily Issue 107 (09/02).