OpenAI's frontier models and Codex workflows will be available through Google Cloud Marketplace, as Google Cloud begins hosting a competitor's models and development workflows. A5X instances powered by NVIDIA Vera Rubin NVL72 will become generally available next month, with networking that supports up to 960,000 Rubin GPUs across sites. Anthropic reportedly posted a net loss of more than $80 billion in the first half of the year; the reported loss and debate over proprietary model pricing bring the cost of enterprise platform expansion into focus. Of 211 tested proteins designed with help from an internal Claude research model, 17 met the criteria for binding scorpion toxins, but their ability to neutralize the toxins in a cellular environment has not yet been confirmed.
OpenAI models and Codex join its distribution channels, while A5X adds next-generation compute resources.
OpenAI and Google Cloud will bring frontier models to Gemini's Model Garden. The partnership also includes Codex workflows, extending beyond model access alone. The models and workflows will be available through Google Cloud Marketplace.
Read original →Google Cloud announced that A5X bare-metal instances powered by NVIDIA Vera Rubin NVL72 will become generally available next month. The accompanying Virgo Network will support clusters of 80,000 GPUs at a single site and up to 960,000 Rubin GPUs across sites. The instances will also support KV cache offloading and reuse across storage tiers, reducing repeated computation in long-context, multi-turn agent tasks and lowering time to first token.
Read original →🔍 Analysis: Adding OpenAI and Codex means Google Cloud does not have to rely solely on its own models to attract workloads. A5X's KV cache offloading and reuse target the cost of repeated computation in long-context, multi-turn agent workloads. For enterprises, model providers and cloud platforms can be selected separately, and procurement comparisons need to account for model capabilities, workflow integration, and inference resources—not just the price of an individual model.
A reported first-half net loss of more than $80 billion puts the business returns from enterprise platform expansion in the spotlight.
Anthropic reportedly posted a net loss of more than $80 billion in the first half of 2026. Some commentators argue that its push into platform features such as source code management and document storage is an effort to become an AI operating system for enterprises and attract customers. Others worry that excessively high proprietary model prices could drive users toward open models or self-hosted alternatives.
Read original →Seventeen designs met binding criteria, but binding and toxin neutralization remain separate validation hurdles.
Of the proteins designed with help from Anthropic's internal Claude research model, 211 were experimentally tested, and 17 met the criteria for binding scorpion neurotoxins. The model worked largely autonomously for 64 hours, with researchers providing guidance through four follow-up prompts. The team will release all tested designs and experimental data. These results establish binding ability only; whether the proteins can neutralize scorpion toxins in a cellular environment remains unconfirmed, so they should not be treated as a finished antivenom.
Read original →NVIDIA, Perplexity, and PyTorch are advancing local coding, on-device actions, and Mac performance, respectively.
At GTC Berlin, NVIDIA announced that its local coding model, Recursive Flint, is now available for its local hardware. The hardware lineup showcased includes RTX Spark, DGX Spark, and DGX Station, targeting Windows personal agents, Linux development, and enterprise workloads, respectively. Other local AI updates announced alongside it include inference acceleration for llama.cpp and vLLM, as well as model checkpoints optimized for local hardware.
Read original →Perplexity Computer can now run entirely locally on a Mac, with the feature available in the Perplexity Mac app. It is not just the model: the agent itself executes on the device and can operate native Mac apps and local files. Tasks involving sensitive data do not require sending that data to the cloud.
Read original →PyTorch 2.15 optimizes model execution on Mac, delivering a 5–6.5× speedup in MoE decoding. Hugging Face's PyTorch team has been developing these changes over the past several months and contributing them upstream to PyTorch. The core approach optimizes the MPS execution path by replacing MPSGraph paths with Metal kernels.
Read original →🔍 Analysis: Recursive Flint addresses the supply of local coding models, Perplexity Computer keeps app and file operations on the Mac, and PyTorch 2.15's 5–6.5× MoE decoding speedup improves performance for local execution. These are not parts of a single product, but each addresses a different factor in making local agents practical. For developers handling sensitive data, evaluating options requires checking not only whether the model runs locally, but also whether tool execution and data flows stay on the device—the latter is what Perplexity is emphasizing here.
PHP2Shell exposes a low-level attack surface, while Claude Mods brings extension code into the user's permission scope.
Wiz disclosed PHP2Shell: vulnerabilities in PHP core that can lead to unauthenticated remote code execution or memory disclosure, with a maximum score of CVSS 9.2. Frontier AI helped researchers develop and adapt exploit chains for multiple applications within hours, without relying on flaws in application code. Affected versions date back to PHP 5.4; PHP released patches on October 21, 2026.
Read original →Claude Code's Mods lets users customize its interface and behavior with JS/TS functions, supporting the CLI, desktop app, and VS Code extension. Unlike configuration hooks that run externally, Mods run inside Claude Code and can redraw the interface and intercept or rewrite tool calls. Mods are not sandboxed: they can read and write files and access keys and sessions with the user's permissions, as well as call models and consume usage.
Read original →🔍 Analysis: PHP2Shell shows how AI can compress exploit-chain development and adaptation into hours. Claude Mods illustrates a different entry point for risk: extensions can access files, keys, and sessions with user permissions without exploiting a vulnerability. The two call for different responses—the former requires patching PHP core, while the latter requires reviewing Mods as executable programs rather than ordinary configuration. Checking application code or prompts alone does not cover either class of risk.