1. Anthropic Partners with SpaceX for Massive Compute Expansion
Anthropic has entered a partnership with SpaceX to utilize the Colossus 1 supercomputer in Memphis, Tennessee. This deal provides Anthropic with over 300 megawatts of compute capacity and access to more than 220,000 NVIDIA GPUs, including H100, H200, and GB200 models. As a result, Anthropic is doubling the five-hour rate limits for Claude Code and increasing API limits for its Opus model. The company is also exploring future collaborations with SpaceX to develop orbital AI compute capacity.
- • Access to 300+ megawatts of compute capacity via Colossus 1.
- • Doubled five-hour rate limits for Claude Code.
- • Increased API rate limits for Claude Opus.
- • Future plans for orbital AI compute capacity.
Increased compute capacity directly translates to higher usage limits and better performance for developers relying on Anthropic's models and agentic tools.
2. US Administration Moves Toward Mandatory AI Safety Vetting
The US AI Safety Institute, rebranded as the Center for AI Standards and Innovation (CAISI), has signed agreements with Google DeepMind, Microsoft, and xAI to conduct pre-deployment safety checks on frontier AI models. The administration is considering an executive order that could mandate government vetting before public release, driven by national security concerns. This policy shift follows recent industry standoffs regarding the potential cybersecurity risks of unreleased models.
- • CAISI established pre-deployment testing agreements with xAI, Microsoft, and Google DeepMind.
- • Potential executive order for mandatory government vetting of frontier models.
- • Focus on national security and cybersecurity risks.
Developers building on frontier models may face new regulatory hurdles and pre-release vetting requirements that could impact deployment timelines.
3. Subquadratic Debuts 12-Million-Token Context Window
Subquadratic has introduced an AI model with a 12-million-token context window, significantly expanding the amount of information a model can process at once. The company claims this model outperforms GPT-5.5 on retrieval benchmarks and has solved the quadratic scaling issues typically associated with long-context attention. Subquadratic plans to release a 50-million-token version in the future.
- • 12-million-token context window.
- • Claims to outperform GPT-5.5 on retrieval benchmarks.
- • Plans for a 50-million-token model.
Massive context windows enable developers to process entire codebases or large document repositories in a single prompt, reducing the need for complex RAG architectures.
4. Anthropic Introduces 'Dreaming' for Managed Agents
Anthropic has introduced a research preview feature called 'dreaming' for its Claude Managed Agents. This process allows agents to review recent events and activity logs to identify and store key information, effectively mitigating context window limitations during lengthy projects. By curating important data between sessions, the feature aims to improve agent performance and memory retention for complex, multi-step workflows.
- • Feature allows agents to review and store key information from logs.
- • Mitigates context window limitations for long-running tasks.
- • Currently available as a research preview for Managed Agents.
This feature provides a native way to manage long-term memory for agents, simplifying the development of complex, multi-step automation workflows.
5. Cloudflare and Stripe Enable Autonomous Agent Provisioning
Cloudflare and Stripe have partnered to allow coding agents to automatically create accounts, manage subscriptions, and register domains. The integration uses a new protocol that enables agents to deploy production applications directly. To prevent unauthorized spending, the system includes a default monthly limit of $100 for agents, and the feature is currently available in open beta via the Stripe CLI.
- • Agents can now provision Cloudflare services and register domains.
- • Integration uses a new protocol for secure payment tokenization.
- • Default $100 monthly spending limit for agents.
This integration removes manual friction from the deployment process, allowing agents to handle infrastructure setup and billing autonomously.
6. CopilotKit Launches Enterprise Platform for Persistent Agent Memory
CopilotKit has launched its Enterprise Intelligence Platform, which provides a managed infrastructure layer for agentic applications. The platform introduces 'Threads' as a core primitive, allowing agents to maintain persistent context across different users, devices, and sessions. It supports various interaction types, including generative UI, voice, and multimodal inputs, and is designed for enterprise environments with support for SOC 2 compliance and air-gapped deployments.
- • Introduces 'Threads' for persistent session memory.
- • Supports generative UI, voice, and multimodal interactions.
- • Enterprise-ready with SOC 2 compliance and air-gapped support.
Persistent memory is a critical requirement for production-grade agents; this platform provides a standardized way to handle state across sessions.
7. Harvey and Artificial Analysis Launch Legal Agent Benchmark
Artificial Analysis has partnered with Harvey to launch the Legal Agent Benchmark (LAB), designed to measure AI performance on complex legal tasks. The benchmark includes 1,200 agentic tasks covering 24 different legal practice areas, providing a standardized way to track and analyze the capabilities of AI agents in the legal sector as of 2026.
- • Benchmark includes 1,200 agentic tasks.
- • Covers 24 distinct legal practice areas.
- • Designed to track performance for 2026-era legal agents.
Standardized benchmarks for domain-specific agents help developers evaluate model suitability for high-stakes professional workflows.