Perplexity released a new orchestrator model in Perplexity Computer, achieving Opus 4.8 grade performance at a lower cost, and the article discusses the value of the harness layer in accumulating knowledge and controlling inference, with the potential to define the next deployment cycle for enterprises. The harness pattern is seen as a key factor in beating general frontier models at specific production tasks.
Kimi K3 ranked second on AA-Briefcase's agentic benchmark, indicating China's frontier labs are competing with Western counterparts on architectural choices for complex multi-step reasoning, and the gap between 'good enough' and state of the art is collapsing rapidly. This development has significant implications for AI deployment timelines.
Kimi K3 has reached the top spot in the DesignArena 3D design rankings with an Elo score of 1450, surpassing Claude 5 Fable and GLM 5.2, and marking a significant improvement over its previous generation. The current top 5 models are Kimi K3, Claude 5 Fable, GLM 5.2, GPT-5.6, and Grok 4.5.
Chamath discusses the rising costs of enterprise AI tokens despite falling unit prices, and Glean's founder reframes the issue around cost per successful task, with a linked piece offering solutions