This week’s AI landscape is marked by massive infrastructure investments, rapid frontier model releases, and evolving agent frameworks that blend openness with enterprise control.
AI Infrastructure & Compute
- SpaceX’s GPU rental deals with Anthropic, Google, and Reflection AI total $2.32 billion per month, annualizing to roughly $28 billion a year—about twice Coreweave’s current revenue 1
- Baseten announced a leaked $13 billion Series F round, underscoring the surge in AI infrastructure financing 1
- OpenAI reported that median internal Codex output tokens surged 56× in Research, 32× in Customer Support, 27× in Engineering, and 13× in Legal since November 2025 2
Frontier Model Releases & Access
- OpenAI announced a restricted preview of its GPT‑5.6 model family—Sol, Terra, Luna—available only to trusted partners; Sol is priced at $5 input/$30 output per 1M tokens and claimed to be its most capable model for coding and cybersecurity tasks 3
- GLM‑5.2, an MIT‑licensed open‑weight model released by Z.ai in mid‑June 2026, quickly matched or exceeded leading closed models such as Claude Opus 4.8 in coding and agent tasks, narrowing the performance gap between U.S. closed labs and Chinese open‑weight alternatives 456
- The open‑artifact ecosystem continues to broaden with releases such as NVIDIA’s Nemotron‑3‑Ultra, Cohere’s Command A+, Zyphra’s ZAYA1‑74B‑preview, and Poolside’s Laguna‑M.1 5
Agent Systems & Meta‑Harnesses
- A step‑by‑step guide shows how to build a fully local coding agent using Ollama to serve open‑weight LLMs like Qwen3.6 35B‑A3B, achieving token generation speeds of roughly 30‑40 tok/sec and offering transparency, cost, and control advantages over proprietary services 7
- Anthropic introduced Claude Tag, a Slack‑native agent that lets teams tag Claude into threads to delegate work asynchronously; internal usage indicates it writes or merges roughly 65% of the product team’s code or PRs in beta for Claude Enterprise and Team plans 8
- Databricks cofounders Matei Zaharia and Reynold Xin unveiled Omnigent, an open‑source meta‑harness providing a common API for agent sessions, security, spend control, and collaboration, alongside LTAP (Lake Transactional/Analytical Processing) as a unified‑storage approach delivering HTAP‑like performance for live operational context 9
- The “Meta‑Harness Summer” highlights open agent efforts such as Qwen‑AgentWorld, OpenThoughts‑Agent, and memory‑centric approaches, while noting strong performance from Chinese open models like GLM‑5.2 and Kimi 6
AI Safety & Red‑Teaming
- Researchers from Gray Swan argue that AI security requires a distinct mindset from traditional cybersecurity, noting that specialized red‑team models like Shade can now outperform humans at breaking models while guardrails such as Cygnal enforce enterprise policies; they warn that a major prompt‑injection breach may be an inevitable gray‑swan event driving growth in AI insurance and compliance 10
Sources
- Latent Space — [AINews] SpaceX is already a $28B/yr Neocloud
- Latent Space — [AINews] OpenAI reports median internal Codex output tokens grew 56x in Research, 32x in Customer Support, 27x in Engineering, and 13x in Legal since November 2025.
- Latent Space — [AINews] OpenAI GPT-5.6 Sol / Terra / Luna — restricted to trusted partners
- Interconnects — GLM-5.2 is the step change for open agents
- Interconnects — Latest open artifacts (#22): Zyphra, Cohere, and Poolside are expanding the breadth of the ecosystem
- Latent Space — [AINews] It's Meta-Harness Summer
- Ahead of AI — Using Local Coding Agents
- Latent Space — [AINews] Claude Tag: Multiplayer, Proactive, Persistent Agents in Slack
- Latent Space — Why the Frontier Ecosystem must be Open — Matei Zaharia and Reynold Xin, Databricks
- Latent Space — Red-Teaming after Mythos — Zico Kolter & Matt Fredrikson, Gray Swan