🌊 West Tide

Thursday, July 9, 2026 ｜ Physix Frontier · West Tide Column

West tide flows east, bringing the cutting edge firsthand.

---

📌 Today's Digest

Today marks a symbolic "unblocking day" for the AI industry. After enduring two weeks of security review by the US government, OpenAI's GPT-5.6 series (three models: Sol / Terra / Luna) has been officially released globally, signaling that frontier model releases have entered a new phase from "technical competition" to "regulatory maneuvering." Meanwhile, Musk announced Grok 4.5 would open to the public on the same day, DeepSeek V4 announced peak-valley pricing, and Google Gemini 3.5 Pro is locked in for July 17—the large model arms race reached white-hot intensity in the second week of July. In autonomous driving, Waymo's data of 500,000 paid rides per week officially solidified the "Year One of Robotaxis," while RoboCup 2026's first-ever full-size humanoid robot football match (11v11) wrote a milestone chapter for embodied intelligence.

---

🤖 AI Large Models (9 items)

🔥 [Headline] OpenAI GPT-5.6 Fully Unblocked: Sol / Terra / Luna Arrow Volley, "Model Family Era" Officially Begins

Importance: ⭐⭐⭐⭐⭐

On July 8, according to an exclusive Axios report confirmed by OpenAI, the US Department of Commerce officially approved the global public release of the GPT-5.6 series models, ending the two-week "limited preview for government-enterprise partners only" phase. OpenAI announced that three models—flagship GPT-5.6 Sol, balanced Terra, and lightweight Luna—would go live fully on July 9, covering ChatGPT, Codex, and the OpenAI API across all channels.

This is the first case in the large model industry where a model was delayed by the government due to security reviews and then "unblocked." GPT-5.6 initially launched a low-key preview on June 26, available only to a few vetted partners. The "Center for AI Standards and Innovation" under the US Department of Commerce directly stationed personnel at OpenAI for security reviews, with core concerns being that the model's superhuman capabilities in cybersecurity offense/defense could be misused.

Clear positioning for the three models:

• GPT-5.6 Sol (Flagship): Specializes in complex reasoning, frontier coding, scientific research, and cybersecurity. Scores 88.8% in standard mode and 91.9% in Ultra mode on the Terminal-Bench 2.1 programming benchmark, surpassing Claude Mythos 5's 88.0%. Introduces Max deep reasoning mode and Ultra multi-sub-Agent collaboration mode. Context window expanded to 1.5 million tokens (a 43% increase over GPT-5.5's 1 million). Pricing: Input $5/million tokens, Output $30/million tokens.
• GPT-5.6 Terra (Balanced): Performance close to GPT-5.5, cost halved. Pricing: $2.5/$15. Industry views it as the "GPT-5.5 Terminator."
• GPT-5.6 Luna (Lightweight): Extreme speed and ultra-low cost, suitable for high-frequency Agent calls. Pricing: $1/$6.

Editor's Short Take: The release of GPT-5.6 marks the shift in AI model competition from a "single flagship" to the era of "family matrices." Sol sets the benchmark, Terra grabs the mid-market, and Luna fights the price war—this combination precisely responds to the pincer attack from Anthropic Fable 5 and DeepSeek V4. More notably, GPT-6 is already in the works; rumors suggest it will completely abandon the old ~4T parameter base, implying the GPT-5.x series may conclude here.

(Source: Axios, OpenAI Official, Xinzhiyuan)

📌 Musk Announces: Grok 4.5 Public Release Same Day, Targeting Opus Level

On July 8, Musk announced on X platform that SpaceX AI (xAI) will open Grok 4.5 to the public on July 9. Musk described it as "an Opus-level model, but faster, more token-efficient, and lower cost." This is another major move by xAI following its acquisition of Cursor's parent company Anysphere, further stirring the competitive landscape of large models.

(Source: X / Twitter, Musk)

📌 Chinese AI Models Account for 30%-46% of Token Usage in US Enterprises, CNBC Deep Dive Sparks Attention

CNBC published a heavy investigation on July 7: Data from the OpenRouter platform shows that Chinese AI models now account for 30%-46% of API token usage among US enterprises (the 12-month average was only 11%). Notably, Zhipu AI's GLM-5.2 saw token volume grow ~27x and customer count grow ~80x in its first week after release, scoring 62.1% on SWE-bench Pro (higher than GPT-5.5's 58.6%), and is under the MIT open-source license. Open-source Chinese models are 60%-90% cheaper than closed-source models from Anthropic/OpenAI, leading enterprises to route routine workloads massively to Chinese models.

(Source: CNBC, OpenRouter, Vercel)

📌 DeepSeek V4 Official Version Announced: Peak-Valley Pricing, Full Launch on July 15

DeepSeek announced the pricing and capability parameters for V4 Official Version on July 8—full launch on July 15. Core highlights: Native 128K context, scalable to 1M; inference throughput improved by ~40% compared to V3, unit token cost down by 30%. It pioneered a "peak-valley pricing" mechanism: prices double during weekday peak hours, while off-peak hours (night + weekend) input prices drop as low as 0.5 RMB/million tokens. This strategy pushes large model commercialization from "grabbing volume" to "refined operations."

(Source: DeepSeek Official, CSDN)

📌 Anthropic Claude 5 Series Fully Ready, Fable 5 Credit Billing Launched

Anthropic's Claude 5 series (Sonnet 5 / Mythos 5 / Fable 5) is fully ready. All models are natively multimodal. Fable 5 credit billing is live ($10/$50 per million tokens); Sonnet 5 offers discounted pricing until August 31 ($2/$10). Additionally, Anthropic published a research paper claiming to have discovered a hidden workspace named "J-Space" inside Claude, which separates background data processing from intent reasoning. Claude Cowork Agent has expanded to mobile and Web.

(Source: Anthropic Official, Tech-Reader)

📋 News Flash

• Google Gemini 3.5 Pro Confirmed for July 17 Release: 2 million token context, native real-time video/audio multimodal, directly targeting GPT-5.6 Sol. Leaked benchmarks show a score of 96.7% on Terminal-Bench 2.1 (SOTA). Simultaneously releasing Gemma 4 12B compact model, natively supporting audio input/output. (Source: Google, Google Blog)
• Beijing Considers Restricting Overseas Access to AI Models: China's Ministry of Commerce convened companies like ByteDance, Alibaba, and Zhipu AI to discuss restricting top-tier models (including Qwen, GLM-5.2) from opening overseas use, potentially involving national security laws. (Source: Reuters)
• DeepSeek Developing Own AI Inference Chips: To reduce dependence on Nvidia and Huawei, while raising $7 billion at a valuation of $52-59 billion. (Source: Reuters)
• UAE TII Releases Falcon H1R 7B: Only 7 billion parameters, scores 88.1% on AIME-24 math benchmark, surpassing Microsoft Phi 4 (14B) and Alibaba Qwen3 (32B), released as open source. (Source: WAM UAE News Agency)
• GPT-6 Development Accelerating: Rumors say OpenAI has decided to completely abandon the old ~4T parameter base "Spud," shifting to a brand-new pre-training architecture, possibly releasing within this month. (Source: Xinzhiyuan)

---

💻 AI Software (8 items)

🔥 [Headline] SpaceX xAI Acquires Cursor Parent Anysphere for $60 Billion—Super Integration Arrives in AI Coding Track

Confirmed by multiple sources, Musk's SpaceX AI (xAI) completed the acquisition of AI coding tool company Anysphere (parent of Cursor) for approximately $60 billion. This is one of the largest M&A deals in the AI sector in H1 2026, marking the shift of the AI coding assistant track from fragmented competition to giant integration.

Cursor is currently the most popular AI-native code editor, valued at ~$9 billion prior to the acquisition. Now under xAI's umbrella, it will deeply integrate Grok model capabilities, forming a complete loop of "model + editor + cloud." Combined with Musk's announcement of Grok 4.5 public release on the same day, xAI is building a full-stack AI coding ecosystem from underlying models to terminal development tools.

Editor's Short Take: The $60 billion acquisition price is both a footnote to the bubble in the AI coding track and a signal flare for giants' "battle for entry points." When OpenAI has Codex, Anthropic has Claude Code, and Google has Gemini Code Assist, xAI chose to directly buy the most successful independent editor—much more efficient than starting from scratch. For developers, whether the Cursor + Grok combo can shake the dominance of VS Code + GitHub Copilot will be a key watch point for the second half of the year.

(Source: Juejin, TechCrunch)

📌 GitHub Copilot Introduces Kimi K2.7 Code: First Open-Weight Model Enters Enterprise-Level Coding Platform

Microsoft announced on July 1 that Kimi K2.7 Code is officially expanding to GitHub Copilot Business and Enterprise plans, becoming the first open-weight model in the Copilot model selector. Developed by Moonshot AI, the model is hosted via Microsoft Azure infrastructure and billed by usage. Enterprise admins must manually enable the model policy; it is disabled by default. This move marks the evolution of enterprise AI coding platforms from "single model" to "model marketplace."

(Source: GitHub Blog, Undercode News)

📌 Alibaba Bans Claude API Completely on July 10, Battle for Domestic Coding Models Takes Off

Alibaba Group internally confirmed that starting July 10, it will completely cut off Claude API and all Anthropic model services, covering Alibaba Cloud Bailian, DingTalk, and all Ant Group business lines, affecting approximately 120,000 developers. Tongyi Lingma urgently launched a "one-click migration from Claude workflow" tool; Alibaba Cloud simultaneously tripled the free quota for Qwen-Coder-Plus to 10 million tokens/day. ByteDance Trae and Tencent CodeBuddy also introduced limited-time migration subsidies. This is the latest landmark event in the decoupling of US-China AI at the developer tool level.

(Source: AI Daily, 36Kr)

📌 Microsoft to Replace Third-Party AI with Proprietary MAI Models, Covering Excel/Outlook etc.

Microsoft announced it will use its proprietary MAI series models to handle routine tasks in everyday applications like Excel and Outlook, aiming to cut compute costs and reduce reliance on third-party labs like OpenAI. Previously, Microsoft launched MAI-Code-1-Flash (GA) and Raptor mini (public preview). This reflects the accelerating trend of tech giants achieving "self-sufficiency" in the AI application layer.

(Source: Enoumen AI Daily, Microsoft)

📋 News Flash

• DoorDash Publishes "DashBench" AI Code Review Test: The Kimi K2.6 + Fable 5 combo captures ~2/3 of historical code issues, with a single review costing only $3.81. AI code review is moving from experiment to practicality. (Source: Enoumen AI Daily)
• Claude Cowork Agent Expands to Mobile: Anthropic extended desktop Agent capabilities to Web and mobile, launching Dispatch for persistent threads and smart routing. Data shows business processes account for 33.4%, content creation 16.4%, and software development only 8.7% of enterprise usage. (Source: Anthropic)
• 83 New Unicorns in US in H1 2026, 51 are AI Companies: Total funding exceeded $52 billion. AI engineering automation firm Prometheus leads with a $41 billion valuation; Recursive achieved a $4.65 billion valuation just 6 months after founding. Nvidia participated in investing in at least 7 new AI unicorns. (Source: TechCrunch, Zhidongxi)
• Four States Sue Meta for $1.4 Trillion: California, Colorado, Kentucky, and New Jersey accuse Meta of designing Facebook/Instagram to be addictive for children. (Source: Reuters)

---

🦾 Humanoid Robots (7 items)

🔥 [Headline] Historic Moment at RoboCup 2026: First 11 vs 11 Full-Size Humanoid Robot Football Match, B-Human Wins 4:0

On July 5, at RoboCup 2026 held in Incheon, South Korea, human robotics history witnessed a landmark moment—two full-size humanoid robot teams completed a full 11 vs 11 football match for the first time. Germany's B-Human team defeated Leipzig's HTWK Robots 4:0, with both teams using humanoid robot platforms designed by Booster Robotics.

This is the closest milestone to the grand goal set since RoboCup's founding in 1997: "Humanoid robots defeat human World Cup champions by 2050." Previously, full-size humanoid robots had never played a complete 11-on-11 match on a real field. The robots demonstrated comprehensive abilities in walking, perception, decision-making, coordination, and competition.

RoboCup Federation Chair Ubbo Visser stated: "This match demonstrates how far humanoid robots have come. New hardware combined with the intelligence brought by AI has pushed humanoid robot football to a whole new level."

Editor's Short Take: While 11v11 looks like sports competition, it is essentially a comprehensive validation of three core technologies: multi-agent collaboration, dynamic balance, and real-time decision-making. Completing this match means humanoid robots have crossed a critical threshold in the closed-loop capability of perception-planning-execution. From industrial sorting to football confrontation, the "generality" of embodied intelligence is evolving at visible speed.

(Source: RoboCup Federation)

📌 Unitree Technology Starts STAR Market IPO Roadshow, Valuation 42 Billion, "First Embodied Stock" Imminent

Following IPO registration approval on June 30, Unitree Technology started its STAR Market IPO roadshow on July 8, planning to raise 4.202 billion RMB with a post-issue valuation of approx 42 billion RMB; institutional subscription oversubscribed ~23x. The company disclosed Q2 2026 shipments of 1,800+ units, with a full-year target of 5,500 units. Funds mainly go towards a production line for 100,000 annual embodied robots and self-developed joint motors/ball screws. This will be the "First Embodied Intelligence Stock" on the A-share market.

(Source: AI Daily, Cailian Press)

📌 MIIT: China's Humanoid Robot Whole-Machine Production Expected to Exceed 100,000 Units in 2026

Gan Xiaobin, Deputy Director-General of the Science and Technology Department at the Ministry of Industry and Information Technology (MIIT), stated at the WAIC 2026 press conference that China's annual humanoid robot whole-machine production is expected to exceed 100,000 units. Currently, the AI application penetration rate among above-scale industrial enterprises exceeds 30%. WAIC 2026 established the first "Embodied Intelligence Pavilion," expecting 100+ humanoid/quadruped/robotic arm robots to exhibit together, with total exhibition area breaking 100,000 square meters for the first time.

(Source: Xinhua)

📌 UBTECH Releases U1 Hyper-Realistic Humanoid Robot: Price up to 990,000 RMB, Pre-orders Exceed 10,000 Units

UBTECH launched the U1 series full-size hyper-realistic humanoid robots at its 2026 Global Launch Event, positioned for daily companionship, emotional support, and social assistance. Male version height 183cm/42kg, female version 168cm/35.2kg, capable of performing >90% of basic human actions. Equipped with the world's first "Emotion Large Model," it can identify 20+ fine-grained emotional states with >90% accuracy. High-end U1 Ultra sells for 990,000 RMB (~$137,000 USD), with pre-orders already exceeding 10,000 units.

(Source: Global Times, People's Daily Overseas Edition)

📋 News Flash

• Apptronik Opens Robot Park Training Facility, Partners with Google DeepMind: Released Apollo 2 humanoid robot (bipedal/wheeled configurations), building a flagship data collection and training center in Austin, Texas. (Source: Robotics & Automation News)
• AGIBOT Rolls Out 15,000th Robot: Milestone for G2 industrial-grade embodied task robots, following nodes at 5,000 and 10,000 units. Also held European Partner Conference in London, featuring the European debut of A3 robot and launch of Robot-as-a-Service model. (Source: Robotics & Automation News)
• Netherlands Opens First Humanoid Robot Innovation Center HAC: Located near Rotterdam, with a 1,000 sqm hands-on testing ground. Chinese companies accounted for >55% (21/38) of humanoid robot exhibitors at CES 2026; HAC is seen as Europe's strategic response. (Source: Robotics Reports)

---

🚗 Autonomous Driving (6 items)

🔥 [Headline] Waymo Leads Comprehensively: Commercial Ops in 8 Cities, Avg 500k Paid Rides/Week, Robotaxi "Really Starting to Make Money"

As of July 2026, Waymo One fully driverless taxi service covers 8 US cities (Phoenix, San Francisco, Los Angeles, Austin, Atlanta, Miami, Las Vegas, Washington D.C.), with a fleet size approaching 4,000 vehicles. According to Whale Factor on July 8, Waymo's weekly paid rides in June averaged 115,000 (YoY growth 130%), adding ~4 million miles of pure driverless mileage per week.

Safety data is striking: Based on 33 million miles of passenger-carrying autonomous driving data, injury-causing collisions caused by Waymo are 85% fewer than human drivers, with zero fatal accidents. Cumulative fully autonomous mileage has surpassed 200 million miles.

Business model becoming clear: ARK Invest estimates Waymo's annualized revenue at ~$180 million, operating costs at ~$400-500 million, but sixth-generation hardware sensor costs are ~50% lower than fifth-gen. In April 2026, Waymo and Uber announced expanded cooperation—Waymo vehicles will accept orders via the Uber App in Austin and Atlanta (starting Q3 2026), directly accessing Uber's 150 million monthly active users.

Editor's Short Take: The inflection point for Waymo from "cash-burning experiment" to "starting to make money" is arriving. ARK predicts operational break-even in mature markets like Phoenix and SF as early as end of 2027, and company-wide profitability by 2029. More importantly, the essence of the Uber partnership solves the biggest non-technical hurdle for Robotaxis—customer acquisition. When 150 million Uber users can hail a driverless car with one click, the "Year One of Robotaxis" truly lives up to its name.

(Source: Whale Factor, Zhiyan Tanxin)

📌 Tesla Cybercab Equipped with Stronger FSD Hardware, Memory Breaks 32GB

According to Notateslaapp, Tesla's soon-to-be-mass-produced Cybercab features an FSD computing platform outperforming all current Model 3/Y family cars, with memory capacity breaking 32GB (current HW4/AI4 uses dual chips totaling 32GB). It may feature an AI4+ pre-production version or dual AI4 motherboard integration; AI5 chips are not expected for mass production until 2027. Each camera module on the Cybercab is equipped with a separate high-pressure cleaning device. Tesla started Cybercab production at the Texas Gigafactory in April this year and began road testing in New Orleans.

(Source: IT Home, Notateslaapp)

📌 Cruise Returns: 300 Vehicles Operating in Four Cities, Targeting Full Driverless in Q4 2026

GM's Cruise, after a comprehensive overhaul following the 2023 pedestrian dragging incident, is returning at a conservative but steady pace. Currently operating ~300 safety-driver-equipped autonomous vehicles in Phoenix, Dallas, Houston, and San Francisco. New CEO Marc Whitten (former Amazon/Microsoft exec) implemented monthly safety updates, planning to transition to full driverless commercial operations in Phoenix first by Q4 2026. The redesigned Origin steering-wheel-less dedicated vehicle has received federal approval, expected to enter limited commercial service in early 2027.

(Source: Whale Factor)

📋 News Flash

• Zoox Publicly Tests Dedicated Vehicle on Las Vegas Strip: Amazon subsidiary, cumulative investment over $6 billion. Bidirectional driving dedicated vehicle design, targeting differentiated market positioning against Waymo. (Source: Whale Factor)
• IIHS Study: Autonomous Ride-Hailing Crash Rate 73% Lower Than Humans: US Insurance Institute for Highway Safety June report showed that controlling for city, time, and road type, police-reported crashes per mile for autonomous ride-hailing were significantly fewer than human-driven ride-hailing. (Source: IIHS)
• NHTSA Still Has 3 Open Investigations into Tesla FSD: As of June 2026, the National Highway Traffic Safety Administration maintains three open investigations related to FSD incidents; Tesla has yet to publish autonomous driving mileage data comparable to Waymo/Cruise. (Source: Whale Factor)

---

🌍 World Models / Physical AI (6 items)

🔥 [Headline] Kyutai Releases MIRA: Playable Rocket League Generated by Pure Neural Network, Zero Game Engine, Zero Physics Code

On July 8, Japanese AI research lab Kyutai, jointly with General Intuition, released the open-source world model MIRA. From 10,000 hours of AI robot battle videos, it generates a fully playable 2v2 Rocket League game relying solely on neural networks—no game engine, no physics engine, no graphics code.

MIRA runs at 20 fps on a single Nvidia GPU, rendering complex game details like boost bars, collision effects, and multi-screen synchronization. Most astonishing (and thought-provoking) is that when goal replays appear, because the model's "memory" is only ~4 seconds, MIRA hallucinates realistic goal footage that never actually happened.

The research team open-sourced all code, training data, and a playable Demo.

Editor's Short Take: MIRA is akin to Odyssey's previous Agora-1 (which generated GoldenEye multiplayer), but Rocket League's physical complexity (ball dynamics, vehicle collisions, aerial control) is far higher than shooting games. Although still far from industrial-grade simulation, it proves one thing: Neural networks can learn sufficiently realistic physical laws from observation and generate interactive worlds in real-time. This is exactly the prototype of what Jensen Huang calls "Physical AI"—not simulation, but emergent physical understanding from data.

(Source: Enoumen AI Daily, Kyutai)

📌 BAAI Releases "Wujie·Physis-v0.1" General World Model: Predicting "Next Physical State"

Chen Boyuan, Head of the Behavior World Model Innovation Center at Beijing Academy of Artificial Intelligence (BAAI), released the Physis-v0.1 general world model at the Beijing News Beike Finance Annual Meeting. Core concept: Shifting from "predicting next Token" to "predicting next physical state." The model is based on latent space physical state modeling, exploring embodied intelligence, industrial simulation, and AI for Science fields, achieving millimeter-level control precision in some scenarios. Chen proposed borrowing the L0-L5 grading approach from autonomous driving to establish a "W0-W5" technical grading system for world models.

(Source: Beijing News Beike Finance)

📌 VicOne Launches Physical AI Security Stress Test at DEF CON 34

Trend Micro subsidiary VicOne announced it will launch the Physical AI Safety Stress Test at DEF CON 34 (Aug 6-9, Las Vegas)—a digital twin CTF challenge testing how cyberattacks alter the perception, decision-making, and physical behavior of AI-driven robots. Covers 6 major security dimensions: prompt injection, AI policy poisoning, adversarial visual inputs, cloud/API security, ROS 2 communication protocol weaknesses, embedded firmware trust chains. VicOne has tracked over 20 public Physical AI security incidents.

(Source: National Law Review, VicOne)

📌 KEPCO Partners with Ndotlight to Build Physical AI Digital Twin

Korea Electric Power Corporation (KEPCO) and 3D AI company Ndotlight launched an 8-month collaborative project. Using their TRINIX platform (based on Neuro-Symbolic architecture, combining 3D CAD engines and generative AI) to automatically create Sim-Ready 3D assets, building digital twins of power facilities in NVIDIA Omniverse and Isaac Sim, verifying the reliability of autonomous robot operation in high-risk environments.

(Source: StartupN)

📋 News Flash

• 36Kr Deep Dive: World Models Becoming AI Industry's Biggest "Concept Basket"—Video generation, autonomous driving, robotics, NVIDIA Omniverse all use the term, but consensus is far from formed. Fei-Fei Li proposed a "Renderer/Simulator/Planner" trichotomy to clarify conceptual boundaries. (Source: 36Kr / GeeTech)
• Waymo Dedicated Vehicle Ojai (Based on Zeekr Platform) Mass Produced: Equipped with 6th Gen Waymo Driver system, first batch of ~100 vehicles deployed for test rides in SF/LA/Phoenix. Marks paradigm shift of Robotaxi from retrofitted cars to dedicated vehicles. (Source: Zhiyan Tanxin)

---

📊 Today's Market Snapshot

• A-shares/HK-stocks AI Concepts Rally Collectively: Zhipu AI up >18%, MINIMAX up nearly 20%, Kingsoft Cloud up >15%, multiple AI concept stocks hit limit-up
• US Pre-market: Alibaba, Kingsoft Cloud up >10%, Baidu up nearly 5%. Jefferies bullish on Alibaba's AI and cloud business prospects
• Galaxy Securities: With improving model capabilities, falling call costs, and maturing multimodal tools, AI applications are expected to accelerate penetration into content production, marketing, e-commerce, gaming, etc.
• Global AI Investment/Funding Hot in H1 2026: Of 83 new US unicorns, 51 are AI companies, with cumulative funding exceeding $52 billion

---

🧠 West Tide · Editor's Perspective

July 9, GPT-5.6 unblocked and Grok 4.5 opened to public on the same day—this is no coincidence, but a microcosm of the white-hot competition in the AI industry.

A structural change worth pondering is occurring: The right to release frontier models is shifting from technical teams to governments. GPT-5.6 and Claude Fable 5 successively went through the three-act drama of "government review-delay-unblock," meaning "stronger model, slower release" may become the norm. For developers and enterprise users, the availability and launch rhythm of AI models will increasingly depend on geopolitical maneuvering rather than pure technical progress.

Meanwhile, "ridiculously cheap" is becoming the new main theme of competition. DeepSeek V4's valley price of 0.5 RMB/million tokens, GPT-5.6 Luna's $1/million tokens, Chinese open-source models capturing 30-46% share in US enterprises—when performance gaps narrow to 5-10 benchmark points, price becomes the decisive variable. Anthropic uses Sonnet 5's $2/$10 discount period to grab market share, OpenAI uses three-tier pricing to cover all scenarios; the ultimate beneficiaries of this price war are every developer calling APIs.

Finally, from "Predicting Tokens" to "Predicting Physical States," AI's battlefield is extending from the bit world to the atom world. World models, Physical AI, embodied intelligence—directions once seen as "future concepts" are turning into reality step-by-step through concrete forms like RoboCup's 11v11 football match, MIRA's playable game, and BAAI's Physis model.

---

📅 Key Points to Watch Next Week

• July 9 (Today): GPT-5.6 Sol / Terra / Luna official full release; Grok 4.5 opens to public
• July 10: Alibaba completely cuts off Claude API, observe effectiveness of domestic coding model takeover
• July 12: Deadline for Anthropic Fable 5 credit billing expansion period
• July 15: DeepSeek V4 full launch
• July 15: Deadline for Anthropic AI for Science grant applications (50 projects, $30k credits each)
• July 17: Google Gemini 3.5 Pro official release (2M context)
• July 17-20: WAIC 2026 World Artificial Intelligence Conference (Shanghai), first Embodied Intelligence Pavilion
• August 1: Deadline for US NSA/CISA frontier model security benchmark delivery—Full deployment of GPT-5.6 and Gemini 3.5 Pro depends on this framework
• August 6-9: DEF CON 34 (Las Vegas), debut of Physical AI Safety Stress Test

---

🌊 West Tide — Overseas AI News Column under Physix Frontier

West tide flows east, bringing the cutting edge firsthand.

© 2026 Physix Frontier · bbs.physixfrontier.com