Unitree's Market Cap Reaches $51 Billion on First Trading Day, Up 460% from IPO Price
Embodied IntelligenceCapital MarketsRobotics
Chinese humanoid robot company Unitree reached a market capitalization of $51 billion on its first trading day, surging 460% above its IPO price, providing the first significant public market valuation reference for the embodied intelligence sector. The company reported RMB 1.7 billion in revenue and RMB 278 million in net profit for 2025, with gross margins exceeding 60%. It delivered over 5,500 humanoid robots last year, reportedly leading globally in delivery scale. This listing enables investors to price and compare humanoid robotics firms based on transparent financials, profitability, and actual shipment volumes.
2
OpenAI Cuts GPT-5.6 Sol Prices by Over 20%, Introduces Spending Controls
Large ModelsAPIPricing
OpenAI announced a price reduction of over 20% for GPT-5.6 Sol's API and Codex credit quotas over the next three months. The降价 has already taken effect for API users and will gradually extend to Codex and ChatGPT Work; subscription usage for Pro, Plus, and Business remains unchanged. The company also introduced new capabilities to track usage and spending per individual API key, allowing organizations or projects to set monthly hard spending limits and programmatically manage budgets via Admin API, reducing the risk of uncontrolled costs when deploying models.
Claude Mythos 5 Security Scanning Goes Public, Backed by $35M Fund
AI SafetyClaudeEnterprise Services
Anthropic has launched a public test of Claude Mythos 5 security scanning for enterprise customers, enabling users to invoke its strongest security capabilities to inspect code without requiring separate access to the model. Suggested patches for identified issues can be generated within the Claude Code web interface, using existing models and billed at standard token rates. The company also announced a $35 million credit fund to support open-source security initiatives, aiming to expand Mythos 5 integration and the Cyber Verification Program, delivering cutting-edge model capabilities to more defensive actors.
Cloudflare Fixes Remote Spectre Variant That Extracts JWT Across Workers
CybersecurityCloud ComputingSide-Channel Attack
Cloudflare disclosed and patched a remote Spectre variant attack on its Workers platform. Researchers were able to extract JWT tokens from co-located Workers at a rate of 12 bits per second, demonstrating that speculative execution side-channel risks exist beyond browsers, extending to multi-tenant serverless environments outside browser contexts. Cloudflare has deployed mitigations in production. The key implication is that attacks can breach tenant isolation and leak authentication materials without traditional memory read privileges, urging platform operators to assess both hardware-level side channels and runtime isolation mechanisms.
5
Google Proposes EnvHarness, Boosting Agent Task Success Rate by 7.12 Percentage Points
AI AgentReinforcement LearningResearch
Google introduced the EnvHarness framework, which improves Agents by enabling task environments to evolve continuously, without modifying underlying environment code. The framework uses Stage to adjust initial states, Contract to rewrite interaction rules, and Chain to sequence or branch multiple environments. Paired with EnvRigger, it follows a four-step closed loop—observe, diagnose, write, and verify—to automatically generate and test environment scaffolds based on Agent trajectories. In SWE-bench Verified experiments, success rates increased from 47.67% to 54.79%, with performance curves showing no plateau even when scaling to 300 environments.
NVIDIA Releases AdaptGrow, Clustering up to One Million Financial Instruments on 16 Nodes
GPU ComputingFinancial AIMachine Learning
NVIDIA released AdaptGrow, a GPU-accelerated clustering solution based on memory-optimized symmetric non-negative matrix factorization. A single GB200 system can handle approximately 100,000 financial instruments, scalable to one million across 16 nodes. Its trace operation reformulation reduces peak memory requirements from ~20n² bytes to ~4n² bytes. The system automatically selects between full-batch or block-wise stochastic updates based on eigenvalue spectrum characteristics. By combining correlation matrices with tail dependence matrices, the method outputs both hard classifications and soft factor loadings, useful for identifying classification shifts, joint crashes, and structural breakpoints.
Slack Launches Slack Code, Integrating AI Coding Agents into Team Collaboration Channels
AI ProgrammingEnterprise CollaborationAI Agent
Slack launched Slack Code, integrating AI coding agents such as Claude Code, Devin, and GitHub Copilot into dedicated code channels. Team members can collaboratively write, review, and deploy code within chat environments while preserving agent execution logs and collaboration history. The feature is now available to all Slack users, marking a shift of coding agents from personal tools into team workflows. Within the ecosystem, NanoClaw further enables the creation of persistent agent teams with unique identities and permissions through a single Slack message, allowing channels to serve simultaneously as task assignment hubs, status synchronization points, and governance interfaces.
8
Claude Code Adds Mobile Initiation of Sessions with Enhanced Cross-Device Real-Time Sync
AI ProgrammingClaude CodeMobile
Claude Code updated its Remote Control feature, allowing users to initiate coding sessions directly from mobile devices. All computers with remote control enabled appear as device cards on the Code tab. Sessions between mobile and desktop are now synchronized in real time; expired sessions are automatically archived upon CLI exit. The system automatically resumes connections if a laptop lid is briefly closed or Wi-Fi switches cause interruptions. The update also synchronizes model and inference intensity settings between mobile and CLI, accelerates loading of large sessions on iOS, and improves usability and execution experience of slash commands on mobile.
Pixel 11 Launches Real-Time Sign Language to Text, Supporting One-Handed and Two-Handed ASL
Accessibility TechnologyComputer VisionOn-Device AI
Google introduced a real-time American Sign Language (ASL) to English text feature for the Pixel 11 series, powered by a multilingual sign language-to-text model that jointly recognizes hand shapes, facial expressions, and body movements. Users can add this feature to the Gboard toolbar and display translations on the external screen by signing with one or two hands in front of the front-facing camera, optionally combining it with Live Transcribe. Google stated the feature was co-developed with the Deaf community to capture natural variations in sign language expression, expanding on-device vision models into everyday communication and accessible input scenarios.