Open-weight models challenge closed models on cost-performance

DeepSeek-V4 Flash delivers 80% of GPT-5.6 Luna's software engineering performance at 1/6 the cost, reinforcing the viability of open-weight models for budget-conscious teams. Muse Spark 1.2 also hits SOTA on Finance Agent v2 at 6.7x cheaper than Opus 5. These benchmarks, alongside Kimi-K3 and Qwen3.8-Max releases, signal a shift toward cost-efficient, specialized models that can run locally or on modest infrastructure.