Anthropic has officially launched Claude Opus 5, now available on platforms including Claude Code, Claude Platform, and Cursor, priced at $5 per million input tokens and $25 per million output tokens—approximately half the cost of Fable 5. The model achieves 43.3% on the Frontier-Bench v0.1 agent coding test (Fable 5 scores 33.7%, more than double Opus 4.8) and leads by 3x over the next-best model on ARC-AGI-3, setting new SOTA across coding, data analysis, design, biology, and knowledge work evaluations. Opus 5 also demonstrates strong resistance to prompt injection attacks, ships with high reasoning mode enabled by default, supports adjustable effort levels, and includes a 2.5x faster mode at 2x base price. New features include mid-conversation tool switching without cache invalidation and fallback routing for safety classification requests. Existing workloads can be migrated via ``/claude-api migrate``. Reviewers praise its cost-effectiveness and autonomous coding capabilities, though note its conversational style is overly cautious and conservative.

OpenAI has launched ChatGPT Health for adult users in the United States, allowing integration with Apple Health and supported medical records to deliver more personalized health conversations. This data will not be used to train foundational models or for ad targeting. Additionally, OpenAI's ChatGPT Work Agent now supports persistent website login, maintaining authenticated sessions across interactions after a single login. While reliability of AI-generated medical advice remains debated, this move marks OpenAI’s deeper expansion into vertical applications.

Reports indicate Stripe is negotiating to acquire AI model routing platform OpenRouter, with a potential valuation as high as $10 billion, although talks could still fall through. OpenRouter provides a unified API aggregating multiple AI models, enabling developers to dynamically select models based on quality, speed, or cost. This aligns with broader industry trends toward model routing—Runway recently launched Media Router, while Cursor and Anthropic are also enhancing their routing capabilities. Stripe is simultaneously pursuing an acquisition of PayPal.

Sony Music has discovered that 61,026 of its sound recordings were used to train AI music company Suno’s models without authorization. The case is being heard in federal court in Boston, centering on whether AI training constitutes 「fair use」, with potential damages possibly reaching $9 billion. A German court in Munich is also set to rule on a similar issue by July 31. Meanwhile, Spotify has removed 75 million AI-generated spam tracks this year, and Deezer receives over 90,000 AI-generated songs daily—more than half of all new uploads—highlighting how generative AI risks diluting creator revenue and increasing copyright infringement. Rights holders can now use audio fingerprinting technology to precisely identify unauthorized content.

The FRONTIER Act is a bipartisan AI regulatory draft establishing a three-tier oversight framework for frontier AI developers: the Base tier (10^26 operations) requires transparency and incident reporting; the Large tier ($50M revenue + $1B investment) adds safety frameworks and audits; the Very Large tier ($50B/$100B thresholds) mandates certification by verified institutions. The bill fixes flaws in the earlier GAAIA proposal by incorporating both revenue and investment metrics. Section 8 authorizes the Secretary of Commerce to suspend a model upon identifying imminent catastrophic risk—via a 45-day provisional order or 90-day final order—but mandatory reporting applies to only about five top companies. Analysts point out major shortcomings, including lack of investigative authority, no severability clause, and absence of whistleblower protections.

Multimodal large model company HiDream.ai has completed three funding rounds within three months, raising over 2.1 billion RMB, pushing its valuation above $1 billion, with participation from national, industrial, and market-driven capital. The company introduced a native multimodal UiT architecture using a single Transformer to process raw signals like pixels and tokens uniformly, avoiding information loss inherent in traditional 「concatenated」 models. Its open-source visual model HiDream-O1 topped the Artificial Analysis text-to-image leaderboard, while the commercial version ranks among the global top three. Its multimodal creation agent product vivago R1 supports infinite-length video generation, achieving an 85% success rate in commercial use and serving over 50 million users worldwide across more than 100 countries.

Hedge fund Bridgewater has unveiled PAT, an internal AI analyst built on LangChain, already deployed to hundreds of investors. It can complete deep research tasks in minutes that would take human experts hours. The system uses permission-aware tools and context, with each investor receiving a customized version. Agreed-upon research plans are converted into dependency-defined data frame tasks, enabling parallel code generation and execution. The team deliberately separates chat and coding layers—investors interact solely in domain language, while coding remains a pure implementation detail. User feedback is transformed into regression tests via a Teach workflow, and suggestions for context or framework improvements are submitted as PRs, forming a continuous learning loop.

A team from Peking University has open-sourced DataFlow-Harness, embedding engineering constraints of data pipelines into a Code Agent to enable natural language requirements to automatically generate inspectable, editable, and reusable platform-native DAG pipelines. The system executes typed changes via an MCP tool layer and undergoes Request-Validate-Commit checks—validating acyclicity and schema compatibility before commit. Experiments show a 93.3% end-to-end pass rate across 12 tasks, reducing cost to $0.261 (down 72.5%) and latency to 95.5 seconds (down 49.9%), outperforming both pure MCP and vanilla Claude Code. Downstream training confirms quality transfer: fine-tuning Qwen2.5-32B on math data pipelines improves AIME24@32 score to 35.9.

NVIDIA's tech blog introduced ModelExpress (MX), which accelerates model weight distribution by intelligently selecting the fastest available source: prioritizing GPU-to-GPU P2P RDMA transfers via NIXL, falling back to object or local storage streaming when peer replicas are unavailable. Two optimizations—pool registration and VMM arena registration—reduce NIXL memory registration overhead by 80–99%. MX extends the transfer mechanism to JIT-compiled kernel caches, allowing replicas to inherit pre-compiled kernels and eliminating redundant warm-up. In testing with DeepSeek-V4 Pro on an 8×B200 node setup, weight transfer takes under 10 seconds, reducing total cold start time from 8 minutes to 1 minute 44 seconds.

TuZhan Intelligence, in collaboration with Peking University and Pengcheng Laboratory, introduced the world model UniWorld-View, which enables novel view synthesis and 4D reconstruction from monocular videos through occlusion-aware point cloud rendering and dual-stream conditional injection. The method uses dual projection to generate occlusion masks, eliminating tearing and light leakage. A geometry stream locks 3D structure while an appearance stream completes details via cross-attention in the Ref-DiT module. A spatiotemporal decoupling strategy first generates static surround views, then produces synchronized multi-view video, with final 4D scenes optimized via 4DGS for free navigation. Full code and weights are open-sourced (PKU-YuanGroup/UniWorld-View) and compatible with domestic Ascend computing platforms.