Artificial Analysis State of AI Q3 2025 is out. https://artificialanalysis.ai/downloads/state-of-ai/2025/Q3-2025-Artificial-Analysis-State-of-AI-Highlights-Report.pdf

Main changes in Q3:

• Two centers of expertise. The US and China are clearly leading by a wide margin.
• Agents are now the core of models. AI is trained to use tools and act autonomously from the start.
• Chats became platforms. They support complex tasks, tools, and integrations — not just Q&A.
• Speech got simpler. New STS models (instead of STT→LLM→TTS) noticeably speed up responses.
• Video generation is mainstream now. High-quality video with audio is in everyday products. Chinese models like Kling 2.5 Turbo are leading.
• OpenAI is open-sourcing again. After a long pause, the company is sharing models. Other open players (e.g., NVIDIA) are improving too.
• No single leader. GPT-5 is strong, but competitors (xAI, Anthropic, Google) are close. Meta is behind. The market is split among several players.
• Inference got cheaper. Using top models now costs about half as much.

What matters for infrastructure teams:

• System simplification. Voice and agents remove extra layers.
• Resource rethinking. Cheap powerful models mean more potential requests or lower spend.