Unitree unveiled UnifoLM-X2-1.0, which it calls the first fully autonomous humanoid powered by a real-time world model — predicting how objects and forces will behave and planning movements on the fly rather than following pre-scripted or teleoperated behavior, demonstrated through autonomous sparring between two robots.

OpenAI began rolling out GPT-6 Astra, disclosing it as the first model to cross the company's 'Critical' internal cybersecurity threshold — capable of finding and exploiting unknown flaws in well-protected systems without step-by-step human guidance. Access is off by default; enterprise admins must manually enable it, with cybersecurity-program partners getting priority access.

Anthropic released Fable 5.1 (general availability) and Mythos 5.1 (restricted-access, the same model with different safeguards for cybersecurity and life-sciences use), cutting cache-read pricing 75% and posting large gains on agentic research and coding benchmarks over Fable 5.

Hyundai Motor Group, Boston Dynamics' majority owner, is considering outside investment as it scales the all-electric Atlas humanoid into mass production, with initial units already committed to Hyundai's Robotics Metaplant and Google DeepMind.

Alibaba open-sourced Qwen3.8-Flash-Next, a 125B-parameter multimodal MoE model with only 6B active parameters per token and a native 262K-token context extendable to 1M, released early to let the community examine the architecture ahead of the full Qwen4 family.

China's Tiangong Ultra, built by the Beijing Humanoid Robot Innovation Center, ran the 100m in 8.86 seconds at the World Humanoid Robot Games — faster than Usain Bolt's 9.58-second human world record — after already breaking its own robot record earlier the same event.

XPeng's robotics division raised over $900 million in its first funding round — the largest private round yet in China's embodied-AI sector — led by IDG Capital with Tencent and Alibaba participating, to push its 76-DoF IRON humanoid toward mass production by the end of 2026.

Unitree Robotics' shares surged as much as 629% on their Shanghai STAR Market trading debut following its $904 million IPO, pushing the company's market value to roughly $66 billion — ahead of Baidu and JD.com — as mainland China's first listed humanoid robot maker.

xAI released Grok 4.6, an upgrade to Grok 4.5 at the same 500K-token context class, aimed at long-running agents, coding, and interactive or visual tasks, keeping the same per-token pricing as its predecessor.

OpenAI said it paused internal work on its unreleased Astra model after safety evaluations couldn't rule out that it had reached a 'critical' cyber-capability threshold — able to find and exploit zero-days in hardened systems without human help — and is now applying stricter isolation, access controls, and monitoring before resuming.

Unitree Robotics priced its Shanghai STAR Market IPO at 150.8 yuan per share, valuing the humanoid robot maker at roughly $9 billion and making it China's first publicly listed humanoid robotics company; the offering — with DeepSeek among the strategic investors — was oversubscribed more than 2,700 times by retail investors.

Demis Hassabis stepped down as Google DeepMind CEO to become chairman and Alphabet chief scientist, handing day-to-day operations to CTO Koray Kavukcuoglu, the same day Alphabet chief scientist Jeff Dean announced he was leaving after 27 years at Google to start Discovery Loop, a startup aimed at automating scientific discovery.

China's large-model API call volume has led the world for fourteen consecutive weeks, coinciding with a sharp rally in domestic AI-themed stock indices.

DeepSeek shipped the official V4-Flash release with stronger autonomous-agent capability and lower API cost, reportedly closing in on Claude Opus on coding benchmarks.

Huawei open-sourced openPangu-2.0-Pro, a ~505B-parameter MoE model trained on Ascend NPUs, releasing weights, inference code, and a technical report.

ByteDance merged its Feishu, Doubao, and Volcano Engine teams into a single AI organization, positioning Doubao as the core product and Volcano Engine as the compute/commercialization layer.

BYD confirmed its first humanoid robot, Xiao Di, will debut in August at its Di Space showrooms — a 1.61-meter, 58.5-kilogram functional prototype meant to greet customers and demonstrate vehicles, with plans for two to three units per store as the company enters the humanoid race against Tesla's Optimus.

Anthropic launched Claude Opus 5, aiming for performance close to its top model at roughly half the price, with a new effort dial to trade off cost and capability.

Figure AI built its 1,000th Figure 03 humanoid at its BotQ factory, up from roughly 350 units in late April, showing its production line can sustain the one-robot-per-hour throughput it first hit three months earlier.

Coverage of Unitree's latest omni model, part of its UnifoLM embodied-intelligence line, letting humanoid robots keep working through interruptions.

OpenAI rolled out the GPT-5.6 series (Sol, Terra, Luna) to the public after clearing an additional US government review process.

xAI launched Grok 4.5, pricing it well below Claude Opus while claiming comparable coding and agent capability.

NVIDIA's own technical blog walks through why the industry is shifting from reactive VLA policies toward World-Action Models that first imagine outcomes, then act.

AgiBot open-sourced the industry's first embodied dataset built specifically for world-model training, deliberately keeping failed and messy real-world interactions rather than only clean demonstrations.

NVIDIA's next-generation Vera Rubin NVL72 platform entered full production, with shipments to OpenAI, CoreWeave, Google Cloud, Microsoft Azure, Meta, and Dell beginning this fall.

Alibaba Cloud launched the Qwen3.7 flagship series at its Cloud Summit, headlined by an all-field thinking mode that unifies reasoning across text, images, and code.

MIT Technology Review's analysis of DeepSeek V4's preview release and why its open-weight strategy keeps reshaping the economics of the model industry.

Google Cloud unveiled its TPU 8t (training) and TPU 8i (inference) chips at Cloud Next, claiming up to 3x faster training and clusters of over a million chips.

Physical Intelligence's steerable robot foundation model, showing a step-change in cross-task and cross-embodiment generalization compared to earlier pi-series models.

A 39-author paper introducing τ₀-VLA, which pairs a memory-augmented high-level planner using world-model-guided test-time search with a generalist low-level policy; trained on 40,000+ hours of real-world robot data, it reaches 45% success on long-horizon manipulation versus 27.5% for a flat, non-hierarchical baseline.

Proposes a dual-system architecture (fast reactive + slow deliberative) for World Action Models, targeting fine-grained manipulation tasks that need both speed and precision.

A tutorial-style paper tracing the conceptual path from pure world models to action-generating World Action Models, aimed at robotics researchers entering the field.

Introduces a regret-aware World-Action Model stack for physical AI systems, letting robots weigh the cost of a wrong action before committing to it.

The first open-source Vision-Language-Action system built for dual-arm, dual-hand manipulation with up to 36 degrees of freedom, using a hybrid teleoperation pipeline (an exoskeleton for arm motion, Apple Vision Pro for markerless hand tracking) and a large embodiment-matched corpus spanning both simulation and real-robot data.

A widely-cited survey formalizing World Action Models (WAMs) as the successor to Vision-Language-Action models, unifying predictive world-state modeling with action generation.

A comprehensive survey mapping how world models are being applied across robot learning, covering architectures, training data, and evaluation methodology.