SpaceX just sealed a potential $6.3 billion deal with Reflection AI to rent out its Project Colossus supercomputers—Nvidia GB300s—to train open-source models. It’s a clear sign that even space-flight companies see data centers and AI compute as a growth area. Meanwhile, OpenAI rolled out GPT-5.5-Cyber in a limited release alongside a revamped Codex Security plugin, a Daybreak partner program and an open-source initiative called Patch the Planet. Instead of giving broad access to GPT-5.5, OpenAI embeds it into existing security products via partners, aiming for tighter controls on who uses what. On the model front, Alibaba’s HappyHorse 1.1 climbed to No. 2 in global AI video model rankings by offering text-to-video, image-to-video and subject-to-video APIs, plus editing tools, under a 40 percent launch discount. Anthropic’s Claude Code still withholds its “Extended Thinking” output—clients only get a summary unless they sign enterprise contracts. Researchers are tackling the gap between smaller and frontier models with “knowledge agents,” specialized LLM wrappers that inject domain data through embedding and multi-pass search. Open models advanced too: GLM-5.2 lands as one of the strongest public releases, though it trails top commercial systems, and forecasts predict model sizes could hit 1.4 quadrillion parameters by 2031 if compute grows as expected. On the tooling side, developers are turning to Codex not just for snippets but as a long-running workspace that retains context across projects, guiding task breakdown and oversight. Moebius, at just 0.22 billion parameters, matches or beats much larger inpainting frameworks with a 15× speed boost. Anthropic plans to start ID checks for flagged Claude users via Persona on July 8, though specifics on who qualifies remain vague. Finally, loop engineering is emerging as a new paradigm: autonomous agents that pick tasks, execute, verify outcomes and iterate, shifting away from one-off prompts to self-driving AI systems.