According to SiliconANGLE, Apple and Google have reached a multi-year agreement to use Gemini series models in the new generation of Siri planned for release later this year, potentially mixing different model sizes per task to optimize cost and latency. The new Siri will be more personalized and capable of executing complex, multi-step tasks across multiple apps. Apple states that Apple Intelligence still runs on-device and in the "private cloud compute" environment, and its integration agreement with OpenAI's ChatGPT remains unchanged; the report also mentions market rumors of an annual fee around $10 billion.

An AI industry briefing stated that Meta has entered into nuclear power supply agreements with Vistra, TerraPower, and Oklo, planning to become the "anchor customer" for new and existing U.S. nuclear power projects to secure power for its AI supercluster Prometheus and long-term compute expansion. The agreements collectively secure about 6.6GW of power, with the first new reactors expected to be delivered as early as 2030 and 2032. This move further elevates "power and grid connectivity" to a core competitive factor in AI infrastructure, alongside supply constraints like GPUs and HBM, forming a barrier to compute scaling.

Nvidia and Lilly announced the establishment of a joint innovation laboratory to accelerate drug discovery using Nvidia's AI and robotics technologies. The companies stated they will invest up to $10 billion over the next five years for talent, infrastructure, and compute power, which includes Vera Rubin chips; the lab will expand Lilly's existing DGX SuperPOD and "AI factory" construction. The two previously collaborated to build a supercomputer for drug discovery, using over 1,000 Blackwell chips. The collaboration will more deeply integrate model training, simulation, and experimental automation, strengthening the closed-loop R&D path of "AI + lab equipment."

Basecamp Research released EDEN, an AI model for "programmable gene insertion," and launched the aiPGI platform, aiming to design enzymes that can perform large-fragment gene insertions at specific sites. The company said the model is trained on its proprietary BaseData™ genomic dataset and accelerated using NVIDIA BioNeMo. Its experiments show successful insertion at over 10,000 disease-related sites; after integrating anti-cancer DNA into T cells at a new "safe harbor" site, lab results showed CAR-T clearance rates exceeding 90%. In an antimicrobial peptide library designed by the same model, 97% of candidates were validated as active, and the company received investment from NVentures.

Google updated the Gemini API's data input capabilities: developers can directly pull content from Google Cloud Storage via registered file URIs, reducing duplicate uploads; it also supports external HTTPS/signed URLs as input, including major storage services like AWS S3 and Azure Blob. For scenarios preferring inline parameter passing, the inline file size limit per request has been increased from 20MB to 100MB (actual size after Base64 encoding varies slightly by type). Google stated this move alleviates production bottlenecks caused by the previous Files API's 48-hour retention limit, benefiting multimodal application scaling and data pipeline simplification.

Google released Gemma Scope 2, an interpretability toolkit for analyzing the internal representations and "reasoning" paths of Gemini 3/Gemma 3 models. Its core involves re-training sparse autoencoders (SAEs) and transcoders for each model layer, with the addition of skip transcoders and cross-layer transcoders to explain multi-step computations and distributed algorithms. Google also provides weights and related components on Hugging Face, positioning the tool for auditing and debugging AI agents, and for researching mitigation strategies for safety issues like jailbreaking, hallucinations, and sycophancy, further advancing "interpretability engineering."

Anthropic launched Claude for Healthcare for medical institutions, emphasizing the provision of "HIPAA-ready" infrastructure to handle workflows involving Protected Health Information (PHI). The product supports secure access to health records via Apple Health and Android Health Connect, stating that related data will not be used for model training; it also integrates data sources like the CMS coverage database, ICD‑10 codes, and PubMed to help verify insurance requirements and pre-authorizations and reduce claim denials. Early adopters include Banner Health and Stanford Health, using it for clinical documentation and regulatory submissions; the company claims some tasks can save 90% of the time, but medical decisions still require professional review.

Adobe announced the integration of OpenAI's GPT‑Image 1.5 image generation model into the Firefly creation platform, continuing to offer third-party capabilities in a "multi-model aggregation" manner. The company provided a phased incentive for subscribers: Firefly Pro and Premium subscribers can use the model for unlimited image generation until January 15. Firefly's billing is primarily point-based: Standard costs $9.99/month for 2000 points, Pro $19.99/month for 4000 points, and Premium $199.99/month for 50,000 points. The platform also integrates Adobe's in-house models, Runway, and Google models, enhancing the end-to-end workflow from design to video editing.

AI Security Operations (SOC) platform Torq announced the completion of a $140 million Series D funding round, reaching a valuation of $1.2 billion. The round was led by Merlin Ventures, with participation from existing investors Bessemer and Insight; the company disclosed total funding raised to date is $332 million. Torq focuses on embedding AI agents into security operations workflows to automatically triage and handle low-confidence alerts, allowing analysts to focus on high-impact threats. The company stated the funds will be used to accelerate platform adoption in commercial and U.S. federal markets and enhance agent capabilities to achieve a higher degree of autonomous security operations.