West Tide Daily | 2026.07.16

🌊 West Tide — Overseas AI News Column under Physix Frontier Bringing the West to the East, cutting-edge first-hand.

Today's Highlights

Today's overseas AI landscape presents three major themes: The model arms race enters an era of "three-tier parallelism" — OpenAI's GPT-5.6 Sol/Terra/Luna trio launches simultaneously, reshaping the pricing system; Google Gemini 3.5 Flash follows closely behind; Thinking Machines shakes up the status quo with a 975B parameter open-source MoE. The humanoid robot sector hits a capital market milestone — Agility Robotics goes public via SPAC at a $2.5 billion valuation, becoming the first listed pure-play humanoid robotics company, while Tesla dismantles its Model S/X production line in 46 days to pave the way for Optimus mass production. Waymo dominates the autonomous driving field, announcing expansion into four cities in a single day, forming a stark contrast with Tesla's mere 20 Robotaxis. Here are today's 35 selected items.

🤖 Large AI Models

1. OpenAI Officially Releases GPT-5.6 Series: Sol/Terra/Luna Trio Launches Simultaneously Summary: On July 9, OpenAI officially launched the three models of the GPT-5.6 series—Sol (flagship deep reasoning, $5/$30/MTok), Terra (balanced general purpose, $2.50/$15), and Luna (lightweight high-speed, $1/$6). Sol achieved an 80-point SOTA on the coding Agent index, with a 1.05M context window shared across the entire series. It is now live on Amazon Bedrock. Source: OpenAI Official / Amazon Bedrock Editor's Note: The three-tier strategy marks OpenAI's shift from a single flagship to refined pricing. Terra offers the best cost-performance ratio and is the top choice for most teams; Luna makes low-cost batch processing possible.

2. SpaceX AI Opens Grok 4.5 Low-Cost Model Summary: Musk's xAI released Grok 4.5 (jointly trained with Cursor), priced at $2/$6 per MTok, focusing on code/legal/financial tasks. Independent evaluations show it has the strongest tool-calling capabilities, but the hallucination rate reaches 54%. Source: xAI Official / Independent Evaluation Editor's Note: One out of every two answers might be fabricated—this is fatal in investment decision-making scenarios. The low-censorship positioning offers differentiation, but EU regulatory risks cannot be ignored.

3. Thinking Machines Releases Inkling: 975B Parameter Open-Source MoE Model Summary: Thinking Machines Lab released Inkling, an MoE architecture with 975B total parameters/41B active parameters, supporting a 1M context window and pre-trained on 45 trillion tokens. A preview of Inkling-Small (12B active parameters) was also released, supporting controllable "thinking intensity" adjustment. Weights are open-sourced on Hugging Face. Source: Thinking Machines Lab Editor's Note: This is currently the largest-scale MoE model in the open-source community. "Controllable thinking intensity" allows developers to flexibly trade off between performance and cost; the combination of open weights + commercial license is highly attractive to the developer ecosystem.

4. OpenAI Removes Usage Caps for GPT-5.6 Sol Summary: OpenAI announced the removal of usage caps for ChatGPT Pro users accessing GPT-5.6 Sol, while Plus users also receive higher daily call quotas. Additionally, Max and Ultra inference control modes were introduced, with Ultra defaulting to coordinating 4 agents for parallel processing. Source: OpenAI Help Center Editor's Note: Removing caps is an aggressive strategy to capture developer mindshare. The multi-agent parallel mode (Ultra) represents a paradigm shift from "single-model inference" to "multi-Agent collaboration."

5. Anthropic Claude Sonnet 5 Goes Live: Agent Workflow Capabilities Catch Up to Opus Summary: Anthropic launched Sonnet 5 with an introductory price of $2/$10 per MTok (valid until August 31), with Agent workflow capabilities directly catching up to Opus 4.8. Note that the new tokenizer may produce 1.0-1.35x token counts for the same text, partially offsetting the actual discount. Source: Anthropic Website / TechCrunch Editor's Note: A classic strategy of using engineering iterations to hedge against flagship bottlenecks. Unit costs for Agent applications continue to fall, benefiting ecosystem scaling. However, ensure you calculate total costs with real samples before migrating.

6. "AI 2040 Plan A" Deceleration Initiative Sparks Industry Debate Summary: Multiple AI researchers jointly issued the "AI 2040 Plan A" initiative, calling for speed bumps on the road to AGI and suggesting the establishment of international AI safety assessment standards. Opponents argue this would hinder innovation. Source: Comprehensive Reporting Editor's Note: The eternal debate of safety vs. innovation heats up again. Notably, this initiative includes insiders from several top labs, giving it different signal significance than before.

7. Google Gemini 3.5 Flash Released: Inference Speed Increased 3x Summary: Google DeepMind released Gemini 3.5 Flash, positioned as a high-speed inference model. While maintaining the core capabilities of the Gemini 3 series, inference speed increased by approximately 3x, and prices dropped by 50%. Previously, Gemini 3.5 Pro was delayed to July 17 due to architectural rebuilding. Source: Google DeepMind / Geeky Gadgets Editor's Note: Google is differentiating between Flash and Pro—Flash focuses on volume, Pro on quality. But the cost of rebuilding Pro is hundreds of millions in GPU resources and loss of core talent.

💻 AI Software

8. NVIDIA Nemotron Open-Source Ecosystem: Inference Costs Reduced 20x Summary: NVIDIA published a series of Nemotron Labs blog posts showcasing multiple enterprises building specialized AI based on Nemotron open-source models. Arcee AI achieved inference costs of ~$0.90/million tokens on the Blackwell platform, about 20x lower than closed-source frontier models. LangChain's Deep Agents achieved the highest Agent accuracy among open models on Nemotron 3 Ultra. Source: NVIDIA Official Blog Editor's Note: From "choosing models" to "customizing models"—Nemotron is becoming the infrastructure layer for enterprise AI. Harvey's legal scenario costs dropping 10x, Glean's search latency reduction—these real cases are more convincing than benchmarks.

9. Anthropic Launches Free Claude for Teachers Education Program Summary: Anthropic announced offering Claude for Teachers free of charge to K-12 educators worldwide, including features like lesson planning, homework grading assistance, and personalized learning path generation. The first phase covers the US and UK. Source: Anthropic Official Editor's Note: Education is one of the socially most valuable scenarios for AI implementation. The free strategy is both brand building and a data flywheel—teacher feedback will help Claude build barriers in the vertical education domain.

10. Meta Muse Image Taken Down Urgently After Three Days Due to Privacy Issues Summary: Meta's Muse Image—the first "Agentic image generation" model capable of calling search and code tools for self-iterative optimization—was urgently taken down just three days after launch due to privacy controversies over training on public Instagram photos by default. Source: Meta Official / The Verge Editor's Note: Training on Instagram data crossed the privacy red line. The direction of "Agent-based image generation" is correct, but data compliance must precede product release; this lesson comes at a significant cost.

11. China MIIT Warns of Backdoor Risks in Claude Code Summary: China's Ministry of Industry and Information Technology (MIIT) issued a security notice pointing out potential backdoor risks in Anthropic's Claude Code tool, advising domestic enterprises to use it cautiously. This is the first time Chinese authorities have issued a security warning specifically targeting overseas AI coding tools. Source: MIIT / 36Kr Editor's Note: The security of AI tools is shifting from a technical issue to a geopolitical one. For enterprises, the selection of code generation tools needs to incorporate supply chain security considerations.

12. AIDE² Achieves Breakthrough in Recursive Self-Improvement Summary: The AI research framework AIDE² released a new version achieving recursive self-improvement capability for models—the model can analyze its own outputs and iteratively optimize, showing a 15-20% improvement in multi-step reasoning tasks. Source: arXiv / Comprehensive Reporting Editor's Note: Self-improvement is a key capability toward AGI. However, it is currently limited to specific task types, still far from general self-evolution. Safety researchers will closely monitor such capabilities.

13. AWS Launches Claude Enterprise Gateway Summary: AWS announced the launch of the Claude Enterprise Gateway on Amazon Bedrock, providing unified API access, usage monitoring, compliance auditing, and cost control features for large enterprise customers. The GPT-5.6 series is also simultaneously live on Bedrock. Source: AWS Official Blog Editor's Note: Enterprise gateways solve governance pain points in large-scale AI deployment. By integrating both OpenAI and Anthropic, AWS strengthens its "AI model supermarket" positioning, giving enterprise customers true multi-cloud, multi-model choices.

14. DCG Releases AI Safety Guardrail Tools Summary: Digital Commerce Global (DCG) released an AI safety guardrail toolkit for enterprises, covering content filtering, output validation, bias detection, and compliance auditing. It can be plug-and-play integrated into mainstream LLMs. Source: Comprehensive Reporting Editor's Note: As AI deploys at scale in enterprises, "guardrails" shift from optional to mandatory. DCG's toolkit targets enterprise compliance pain points, capturing the market timing precisely.

🦾 Humanoid Robots

15. Agility Robotics Goes Public via SPAC at $2.5B Valuation—First Listed Pure-Play Humanoid Robot Company Summary: Agility Robotics announced going public through a SPAC merger with Churchill Capital Corp XI, valued at $2.5 billion, raising $620 million+. Digit is already commercially operating at 9 customer sites including Amazon, GXO, and Toyota, accumulating over 65,000 hours. It will trade on Nasdaq as "AGLT". Source: TechCrunch / The Robot Report Editor's Note: Humanoid robots moving from venture bets to public market tickers is a significant milestone. However, the SPAC model itself carries volatility risks, and the company did not disclose revenue—$300M is order backlog, not confirmed revenue.

16. Tesla Dismantles Model S/X Line in 46 Days to Pave Way for Optimus Mass Production Summary: Tesla's Fremont factory completed the dismantling of the Model S/X production line in just 46 days; the site will be converted to Optimus humanoid robot manufacturing. The last Model S/X rolled off the line on May 10. Optimus Gen 3 is expected to start mass production in July-August, with a target annual capacity of one million units. Source: IT Home / Electrek Editor's Note: Switching strategic identity from flagship EVs to robot workshops in 46 days. But Figure has delivered 350+ units running in real factories, while Optimus is still "about to be unveiled"—capacity plans look strong on paper, but shipments > PPT.

17. Boston Dynamics Atlas Enters Actual Deployment in Hyundai Motor Factories Summary: Hyundai plans to deploy Atlas humanoid robots for parts sorting in its Georgia, USA factory by 2028, with a long-term goal of reaching 30,000 units annual production by 2030. The electric Atlas has 56 degrees of freedom and a 30kg payload. All 2026 capacity is locked by Hyundai and Google DeepMind. Source: Boston Dynamics / Korean Financial News Editor's Note: The best robots serve their own factories first—this is a verification loop and a necessary step before external supply. The technical ceiling of 56 DOF is unmatched, but "unavailable for purchase" is a minus for ecosystem building.

18. Figure 03 Enters BMW Factory: Over 350 Units Delivered Summary: Figure's 03-type robots have delivered over 350 units to partners, achieving autonomous parts sorting and loading operations at BMW's Spartanburg plant. In May 2026, they completed 200 hours of continuous autonomous package sorting with zero human intervention. Production line throughput has increased from 1 unit/day to 1 unit/hour. Source: Frontier News / The Robot Report Editor's Note: 350+ deliveries + 200 hours of continuous operation—Figure has established the industry's hardest credit endorsement with data. The Helix 02 VLA model has been validated in real production environments.

19. 1X NEO Pre-orders Sold Out in 5 Days: $499/Month Subscription Model Summary: 1X NEO humanoid robot is priced at $20,000 or a $499/month subscription. All 10,000 units of annual capacity sold out within 5 days of pre-order opening, with deliveries expected to begin late 2026. It features a 25-DOF tendon-driven hand system. Source: Forbes Editor's Note: Demand signals for consumer humanoid robots are stronger than expected. The $499/month subscription lowers the trial barrier—if delivery experience meets standards, it could become the "iPhone moment" for the robotics industry.

20. Boston Dynamics Launches Three Commercial Robots Simultaneously Summary: Boston Dynamics simultaneously launched the industrial Atlas, upgraded Spot inspection robot, and Stretch warehouse robot, forming a complete product matrix covering industry, inspection, and warehousing. Source: Comprehensive Reporting Editor's Note: Product matrixing is a necessary path to commercialization. Single products struggle to support independent business entities; multi-scenario coverage maximizes platform value.

21. NVIDIA BioNeMo: AI Agent Platform for Life Sciences Summary: NVIDIA launched the BioNeMo life sciences Agent platform, integrating protein structure prediction, drug molecule design, and genomic analysis capabilities, providing end-to-end AI R&D solutions for biopharmaceutical companies. Source: NVIDIA Official Editor's Note: From chips to models to Agents—NVIDIA's layout in life sciences is becoming increasingly complete. The commercialization path of AI+biomedicine is clearer than AI+manufacturing because data is more standardized.

🚗 Autonomous Driving

22. Waymo Announces Four-City Expansion in One Day: Fleet Size Reaches 3,500 Vehicles Summary: Waymo announced expanding Robotaxi service coverage to four cities on the same day, bringing its fleet size to approximately 3,500 vehicles. The goal is to complete 1 million paid rides weekly. A dedicated Ojai vehicle model was also introduced. Source: Comprehensive Industry Reporting Editor's Note: 3,500 vehicles vs. Tesla's 20—Waymo's lead in the Robotaxi track is orders of magnitude. The execution of a four-city expansion in one day shows Waymo has moved from "can we do it?" to "how fast can we do it?"

23. Tesla Robotaxi Operates Only 20 Vehicles: Gap with Waymo Becomes Obvious Summary: Tesla's Robotaxi service in Austin currently operates only about 20 vehicles, forming a stark contrast with Waymo's 3,500. Safety data shows only 1 accident since mid-April (human rear-end collision), with no liability attributed to the autonomous system. Source: Phoenix Net / IT Home Editor's Note: The gap between 20 and 3,500 vehicles is not just about scale, but a divergence in commercialization paths. Tesla relies on FSD mass production data to catch up, while Waymo takes the dedicated fleet route. Which path will succeed first?

24. Waymo Safety Data: Fatal Accident Rate Reduced by 92% Summary: Waymo released latest safety data showing that under supervised autonomous driving mode, its Robotaxi fatal accident rate is reduced by 92% compared to human driving, and serious injury accident rate is reduced by 75%. Cumulative autonomous driving mileage exceeds hundreds of millions of miles. Source: Waymo Official Editor's Note: A 92% reduction in fatal accidents is a very strong safety argument. But there remains a huge gap between public perception and actual data—every accident is still scrutinized magnifiedly.

25. Waymo Unveils Dedicated Ojai Robotaxi Model Summary: Waymo showcased the Ojai dedicated Robotaxi model developed in cooperation with Geely, removing the steering wheel and pedals, optimized for urban mobility scenarios. Large-scale deployment is planned for 2027. Source: Comprehensive Industry Reporting Editor's Note: Dedicated models signify a shift from "modifying existing cars" to "born for Robotaxi." Removing the steering wheel is the ultimate vote of confidence in autonomous driving technology.

26. Waymo Internationalization: London and Tokyo Enter Testing Phase Summary: Waymo announced it will start Robotaxi testing in London and Tokyo, marking its first entry into markets outside the US. They are working with local partners to advance regulatory adaptation and road testing. Source: Comprehensive Industry Reporting Editor's Note: Internationalization is a necessary path for Robotaxi commercialization. The choice of London and Tokyo is deliberate—one is a global financial center with right-hand drive, the other a model of high-density urban living.

27. Uber and Waymo Achieve Interoperability: Robotaxis Can Be Hailed via Uber Summary: Uber and Waymo reached a partnership allowing users to hail Waymo Robotaxis directly through the Uber App. The two companies are advancing system interoperability, expected to go live in multiple cities by year-end. Source: Comprehensive Industry Reporting Editor's Note: From competition to cooperation—Uber chose a "platformization" strategy, incorporating Robotaxis into its existing ride-hailing network. This is a massive distribution channel for Waymo and a future-proof layout for Uber.

28. Discussion Arises on Whether Mercedes S-Class L3 Autonomous Driving Can Be Used in China Summary: Mercedes' Drive Pilot system has received certification in Germany/US, but lacks legal conditions for use in China. The fragmentation of global L3 regulations is intensifying. Source: Sina Auto Editor's Note: Technical leadership does not equal market leadership. Global regulatory fragmentation is one of the biggest obstacles to autonomous driving commercialization; the advancement of ADS GTR unified regulations will change this landscape.

🌐 World Models/Physical AI

29. Xiaomi Releases and Open-Sources Xiaomi-Robotics-U0: 38B Parameter Unified Embodied Generation Model Summary: Xiaomi released U0, a 38-billion-parameter multimodal autoregressive world foundation model, unifying embodied scene generation, embodied transfer, interactive video generation, and general image generation into a single framework for the first time. It ranked first among 126 models in WorldArena evaluation, improving task completion progress in real-machine tests by 26.3%. The FlashAR+ inference acceleration scheme improved generation efficiency by 82.9x. Source: Xiaomi Robotics Division / Zhidongxi Editor's Note: From "generating data separately" to "one model handling four types of tasks"—U0 provides a unified base for scalable production of robot training data. The open-source + commercial license strategy helps quickly establish an ecosystem.

30. Ant Group Releases LingBot-World: World Model for 1-Hour Coherent Video Simulation Summary: Ant Group released the LingBot Brain Full Stack 2.0 series, where LingBot-World evolves video generation models into interactive world simulators, supporting 720p/60fps HD real-time interaction and generating 1-hour coherent simulation videos continuously. It supports button controls and natural language to change world states in real time. Source: Ant Group Official Editor's Note: From "video generation" to "interactive world simulation"—this is a key leap for world models from ornamental to practical. The real-time nature of 720p/60fps makes it truly usable for robot policy training.

31. Landscape of Three Major Schools of World Models Forms Summary: Currently, three major technical schools have formed in the world model field: the video generation school (Sora/Seedance etc., focusing on visual fidelity), the physical simulation school (Cosmos 3/LingBot-World etc., focusing on physical accuracy), and the unified model school (U0 etc., fusing multimodal capabilities). Each school has advantages in different scenarios. Source: Comprehensive Industry Analysis Editor's Note: The three schools are not mutually exclusive—they may ultimately converge. Visual quality from the video generation school + accuracy from the physical simulation school + flexibility from the unified model school constitute the ideal form of a world model.

32. VLA Becomes Core Paradigm for Embodied Intelligence Summary: Vision-Language-Action (VLA) models have become the core technical paradigm for embodied intelligence in 2026. XPeng VLA 2.0 achieved 50% mileage autonomous driving in AVs, Figure Helix 02 validated VLA's industrial application capabilities in factories, and GR00T 1.7 provided a standardized VLA foundation model for the industry. Source: Comprehensive Industry Analysis Editor's Note: VLA is the bridge connecting "seeing," "understanding," and "doing." From autonomous driving to humanoid robots, VLA is becoming the universal language of embodied intelligence.

33. World Labs Releases Marble: 3D-Native World Model Summary: World Labs (founded by Fei-Fei Li) released Marble, the first 3D-native world model. Unlike 2D sequences of video generation, Marble directly models and reasons in 3D space, supporting 6DoF free viewpoint movement and physical interactions. Source: World Labs Official Editor's Note: From 2D video to 3D space—this is a qualitative dimensional change for world models. The academic accumulation of Fei-Fei Li's team is translating into product strength; the 3D-native route has natural advantages in robotics and AR/VR scenarios.

34. Real2Sim Data Engine: Bridge from Real World to Simulation World Summary: Multiple institutions jointly released the Real2Sim data engine, which can automatically convert real-world video and sensor data into interactive simulation environments, providing high-fidelity training data for robots and autonomous driving. Source: arXiv / Comprehensive Reporting Editor's Note: The inverse process of Sim2Real (simulation to reality)—Real2Sim makes "learning from the real world" more efficient. When simulation environments can be automatically constructed from real data, the data bottleneck is significantly alleviated.

35. NVIDIA Teams Up with HuggingFace to Advance LeRobot Open-Source Robot Platform Summary: NVIDIA and HuggingFace deepened cooperation, integrating Cosmos 3 physical AI models and Isaac Sim simulation environments into the LeRobot open-source platform, lowering the barrier for robot development. Fine-tuning and deployment of the GR00T 1.7 model are already supported. Source: NVIDIA / HuggingFace Editor's Note: A strong alliance in the open-source ecosystem. NVIDIA provides compute and simulation, HuggingFace provides community and toolchains—LeRobot is poised to become the "Transformers library" for robotics.