Model Competition Intensifies with Cost-Quality Trade-offs

Alibaba's Qwen3.8-Max beats GPT-5.6 Sol on multiple benchmarks; Qwen3.8-27B launched. DeepSeek plans price increase. Cost optimization strategies: routing (DeepSeek V4 Flash 6x cheaper than Luna), Capy v2 claims 50% lower cost, BDH-CQ latent reasoning at $0.0007/task. A 27B agent beat Claude Opus 4.8 and GPT-5.5 on research replication via RL. Gemini 3.7 Flash shows strong throughput. GPT-5.6 Sol proved 22-year-old math conjecture. GLM-5.3 nearly matches Sol and Fable on coding/cyber. New benchmarks: SWE-Bench ProMax, Evo-Bench. Research: Ornith 9B beats 31B+ models, Skaling scaling laws, LycheeMemory V2, Alaya-EVOKE, PlayWorld, LATTICE tabular model with graph priors. Hugging Face reports small models dominate usage. Intern-S2-Mobius decouples knowledge and reasoning, achieving 4x inference speedup on 35B models. Recent: Synthefy launched foundation model for structured numerical data. GenRouter achieves 95% cost reduction in agentic image generation. UI-Mate advances open-weight GUI agents with 77% on OSWorld-Verified. Gemini 3.7 Flash praised for speed. GLM-5.3 vs Fable 5 cost comparison shows 15x reduction. ClawGym II explores black-box RL on agent harnesses. VibeWorlding shows multimodal agents constructing 3D worlds. New: FreeToken enables edge-native MoE serving for 753B models on single GPU; AxiomProver formalized BGP246 theorem; HumanEvals library for human evaluation; Agentic ESOpt fine-tunes long-horizon agents with minimal GPU; Box integrates Gemini Multimodal Embeddings 2 for enterprise agents; Capability-driven data design for generalist image generation. Latest: GEN-1.5 embodied foundation models achieve one-shot in-context learning for physical tasks. Foundation AI's Antares 3B rivals GPT-5.5 on vulnerability localization. Looped language models improve compositional tool calling. Higgsfield raised $400M Series B with Together AI for video generation inference. New today: DeepSeek launched V4-Flash-Vision-Exp test model targeting Anthropic's Opus 4.8 on multimodal agentic tasks. DeepSeek Flash adoption surging 10-100x in enterprises. Nvidia AVO scores 100% on ARC-AGI-3 public set (40% private, harness on Claude Opus 5). Chain-of-Experience improves LLMs 5.6% with 19% cost reduction. Tabular foundation model primer highlights domain-specific models outperform LLMs on structured data. Google released TabFM, a tabular foundation model with in-context learning, though GBDT still dominates. Ornith-1.5 (9B) matches Claude Opus 4.8 on reasoning/agentic benchmarks via self-improvement training. Joon Sung Park argues simulation, not reasoning, is the path to human-like agents, with 85% accuracy on digital twins. FlashPrefill V2 achieves 47x speedup for long-context serving, further driving cost reduction.