Frontier Model Competition and Agent Trends Accelerate

OpenAI crushed humans at AWTF. GPT-5.6 Luna outperforms GPT-5.5 at 10% cost; GPT-5.6 now default in Microsoft 365 Copilot. Kimi K3 (2.8T params, 1M context, native multimodal) tops web engineering benchmark ahead of Fable, but Bindureddy notes speed is the only issue. Google releases Gemini 3.6 Flash, 3.5 Flash-Lite, 3.5 Flash Cyber but no 3.5 Pro; Bindureddy flags Gemini 3.6 Flash regressing below 3.5 Flash, worse performance and higher cost than Grok/Luna. Sakana AI's Picbreeder experiment explores open-ended creativity with VLMs (GECCO 2026). Muse Spark 1.1 outperforms Opus 4.8 and Grok 4.5 on OOD evals. New benchmarks: CausalDS, Harbor-Index, Long-Horizon-Terminal-Bench (best model 15.2% pass@1). New methods: UP, Jet-Long, Flash-BoN, loop engineering for autonomous ML research. Ethan Mollick warns AI strategies from late 2025 are obsolete. SearchOS-V1, SEED, RecGPT-V3, S1-Omni, Cura 1T, RESOURCE2SKILL, DSWorld, RAGU. US threatens sanctions on Chinese AI models over IP theft. Arcee pushes back on narrative that Chinese open-weight models are dangerous. New: Self-improving agent harness using static analysis and LLM-assisted mapping improves coding agent win rates by 10-17 points. US sanctions on Moonshot AI moving from threat to action. Ethan Mollick clarified GPT-5.6 Pro (chatbot) is smarter than Codex Ultra. Google Gemini usage data shows multimodal AI useful for manual labor. Ethan Mollick's joke prompt led Codex to autonomously create BenchBenchBenchBenchBench and write a paper, demonstrating growing autonomous research capability.