GPT-5.6 Sol Ultra Released; Kimi K3 Officially Confirmed; Frontier Math Results

GPT-5.6 Sol Ultra released to Codex; Sol is A-tier coder (56/100) but not Fable-level (91/100). Luna matches 5.5 at 10% cost. Fable 5 restored after government intervention; performance degradation reported. @bindureddy prefers Fable 5. Roboflow benchmarks show Sol's detection mAP jumped from 13.8 to 46.2. Kimi K3 now officially released: 2.8T params, 1M context, beats Sol and Fable in spatial reasoning and 3D generation. +38% over K2.7 on DeepSWE, #3 overall as first open-weights frontier model. GPT-5.6 Pro used a prompt to close a 30-year convex optimization gap and disproved the Dinitz-Garg-Goemans conjecture. Separately, Grok 4.5 also disproved a 30-year-old graph theory conjecture. K3 autonomously optimized GPU kernels in a 15-hour run, halving compute time. @julien_c suggests K3 > GLM 5.2. Practical: open weights not accessible—hardware costs high; K3 can be used as teacher for distillation. Fable 5 reportedly helped produce a potential Jacobian conjecture counterexample. K3 and Fable show higher agreement on DeepSWE than with Sol, suggesting reasoning convergence. Confirmed: Moonshot AI distilled Fable for K3. New: Kimi K3 scaling law tech report shows ~2.5x improvement over K2; weights now downloadable on Hugging Face. New: Two Korean 700B MoE models released within 48 hours—A.X K2 (688B, FP8-trained, 97.1 AIME26) and K-EXAONE 2.0 (750B, upcycled, 262K context, strong long-context retrieval). Both Apache 2.0, highlighting post-training tradeoffs. New: Alibaba unveiled Qwen3.8-Max (2.4T params), an open-weight coding model for agentic workflows.