Ant Group's Bailing team just dropped Ring-2.6-1T, a trillion-parameter reasoning model (with 63 billion activated parameters) that's already outperforming GPT-5.4 and Gemini 3.1 Pro on one key benchmark. The model uses a new "dynamic thinking intensity" mechanism to balance cognitive depth, token cost, and execution speed.
https://twitter.com/AntLingAGI/status/2052808934390661134
Depending on the workload, Ring offers two modes: high and xhigh. In high mode — designed for agentic tasks like multi-step execution and tool calling — it scored 87.60 on PinchBench, ahead of GPT-5.4 xHigh and Gemini-3.1-Pro high. It also earned 63.82 on ClawEval. In xhigh mode (deep thinking for math and science), it hit 95.83 on AIME 26 and 88.27 on GPQA Diamond.
Ant says the mechanism helps reduce token waste by scaling compute to the task — a text format conversion doesn't need the same firepower as a math competition. The aim is to make Ring a go-to backbone for high-frequency agent scenarios like tool orchestration, coding, and multi-turn interactions. It joins a wave of new models from Chinese AI labs — Zhipu AI, for instance, recently released its multimodal programming model GLM-5V-Turbo. Starting today, you can try Ring for free on OpenRouter through Novita — for one week, ending May 15. Weights will be open-sourced soon.