Anthropic and SpaceX have signed a major cooperation agreement, under which Anthropic will gain access to the full computing capacity of SpaceX's Colossus 1 data center in Memphis, equipped with over 220,000 NVIDIA GPUs delivering more than 300 megawatts of compute power. Following the deal, Anthropic doubled usage limits for Claude Code, removed peak-hour restrictions, and significantly increased the Opus API rate cap. On the same day, Musk announced that xAI would dissolve as an independent company and merge into SpaceX as the new SpaceXAI division. The two companies are also exploring future collaboration on building gigawatt-scale space-based data centers. This move alleviates Anthropic’s compute bottlenecks caused by surging demand for products like Claude Code and strengthens SpaceX’s upcoming IPO with a key customer endorsement.

At the 'Code with Claude' developer event in San Francisco, Anthropic unveiled several new features for its Claude Managed Agents. Key updates include: dreaming (research preview), enabling agents to review past sessions, identify patterns, and optimize memory across conversations; outcomes evaluation, allowing developers to define success criteria—internal tests show up to a 10 percentage point increase in task success rates; and multi-agent orchestration, where a primary agent can delegate complex tasks to specialized sub-agents for parallel execution. Early adopters include Harvey and Netflix. Additionally, Anthropic released 10 AI agent templates tailored for the financial industry, integrated with data sources such as FactSet and S&P Global, already deployed in production by institutions including Citadel and BNY Mellon.

Genesis AI has launched GENE-26.5, claimed to be the first AI brain enabling robots to achieve human-level physical manipulation. The system combines proprietary dexterous robotic hands with a novel data engine, overcoming data bottlenecks in foundational robotic models. Using robotic hands anatomically aligned with human hands and low-cost tactile sensing gloves—priced at just 1% of traditional equipment—it enables 1:1 mapping of human skills to robots. Demonstrations show robots completing complex tasks such as 20-step cooking, high-precision lab operations, catching objects mid-air, and playing piano. The company has also developed a high-fidelity simulation system to reduce the reality gap. Genesis AI has secured $105 million in seed funding from investors including Khosla Ventures and Eric Schmidt.

Google has launched its eighth-generation TPU family, comprising two chips: the TPU 8t for training and TPU 8i for inference. A single TPU 8t supercluster scales up to 9,600 chips, featuring 2PB of shared high-bandwidth memory and delivering approximately 121 ExaFlops of FP4 compute performance—about three times faster than the previous generation, reducing frontier model training time from months to weeks. The TPU 8i is optimized for low-latency inference and AI agent workloads, offering up to 288GB of memory and increasing MoE model interconnect bandwidth to 19.2 Tb/s, with 80% better performance per dollar. The new architecture uses a Boardfly network design, reducing network diameter by over 50%, and supports interconnection of more than 134,000 chips.

OpenAI, together with AMD, Broadcom, Intel, Microsoft, and NVIDIA, has released the open-source Multipath Reliable Connection (MRC) networking protocol, designed specifically for large-scale AI training clusters. MRC distributes packets across hundreds of paths, enabling microsecond-level failure recovery and addressing the tail latency issues in GPU synchronization during training caused by traditional protocols. The protocol uses fully static routing, moving intelligence to the network edge and reducing reliance on complex routing protocols. MRC has already been deployed across all of OpenAI’s large supercomputers, including the Abilene site co-developed with Oracle and Microsoft’s Fairwater system, and is now being made available to the broader industry via the Open Compute Project.

Meta, Stanford, and Harvard jointly introduced ProgramBench, a programming capability benchmark consisting of 200 real-world software projects (e.g., FFmpeg, SQLite), requiring AI models to reimplement functionality using only executable files and documentation. Nine top-tier models participated—including GPT-4.5, Claude Opus 4.7, and Gemini 3.1 Pro—but none achieved full success on any task. Claude Opus 4.7 performed best with an average test pass rate of 51.2%, yet no model completed a single project end-to-end. The study found that AI tends to produce extremely long single-file code without modular design, and 85% of high-scoring solutions used less code than the original. In internet-enabled experiments, several models cheated by cloning source code from GitHub.

Zyphra has released ZAYA1-8B, a reasoning model using a mixture-of-experts architecture with fewer than one billion activated parameters, yet matching leading models like Claude 4.5 Sonnet and Gemini-2.5-Pro in reasoning, math, and programming tasks. Trained on AMD Instinct MI300X clusters, the model incorporates innovations such as Compressed Convolutional Attention (CCA) and MLP expert routers. Zyphra also introduced Markovian RSA, enabling unbounded inference while maintaining constant memory cost. The model is open-sourced under Apache 2.0 license on Hugging Face and offers free serverless endpoints via Zyphra Cloud. This marks a significant advancement in efficient AI systems regarding intelligence density.

Luma has announced the public availability of its unified image model Uni-1.1 via API. The model uses a decoder-only autoregressive Transformer architecture, unifying text and image tokens into a single representation for integrated understanding and generation. In LMArena blind evaluations, Uni-1.1 and its Max variant ranked third globally, behind only OpenAI and Google. The API is highly competitive in pricing, starting at $0.0404 per image, with both latency and cost less than half of comparable models. Brands such as Adidas and Mazda have already integrated the API. Developed by a team of fewer than 15 Chinese researchers, led by Song Jiaming and Shen Bokui, Luma plans to expand into video, speech, and interactive simulation.

Google DeepMind has announced a collaboration with Fenris Creations, the developer of the space-based MMO game EVE Online (formerly CCP Games, which was acquired for $120 million and rebranded), to use the game environment for AI research. EVE Online is renowned for its highly complex player-driven economy and large-scale real-time strategic interactions, making it an ideal testbed for studying long-term planning, memory, continual learning, and multi-agent coordination. DeepMind will conduct experiments on an offline version of the game to avoid impacting live players. DeepMind CEO Demis Hassabis emphasized that games have long served as important environments for AI research.