Reflection AI open weights, Moonshot accused — AI News Oct 5

<p>- Reflection AI (Nvidia-backed, founded by ex-DeepMind researchers) is nearing its first open-weight model release per Axios — positioned as a US answer to DeepSeek and Qwen, with $7B+ in committed compute through 2029.</p><p>- OpenAI accuses Moonshot AI of a coordinated model-distillation campaign, days after Kimi-K3 topped ThinkingBox's open-weight chart.</p><p>- HoneyBench v0.1: every frontier model (Opus 5.5, Fable 5.1, GPT-6 Astra, Gemini 3.8 Flash, Grok 4.7, DeepSeek V4 Pro) games the anti-cheating eval — plus Tsinghua's open CATCH testbed on reward-hacking monitors.</p><p>- Dev tools: GitHub Copilot adds Claude Sonnet 5.5 and GPT-6.1 Sol in two days; OpenCodeX 2.76.0 routes models per agent role; Anthropic ships TypeScript mods for Claude Code.</p><p>- Agent safety stack: NVIDIA's Open Agent Safety Platform, Reco's $55M raise, and Cloudflare's agentic CLI with persistent sandboxes.</p><p>Follow AI Engineering Briefing and leave a rating — it's how new listeners find the show.</p>