Ant Group Launches Trillion-Parameter Model Ring-2.6-1T: PinchBench Scored 87.60, Outperforming GPT-5.4

According to Sentinel AI monitoring, Ant Group's Phoenix Mega Model Team has launched the trillion-parameter flagship deliberation model Ring-2.6-1T (with 630 billion activated parameters). This model is specifically designed for complex tasks and production environments, with a core addition of the "Dynamic Deliberation Intensity" mechanism, allowing the system to flexibly balance cognitive depth, token cost, and execution speed.

Based on varying computational load requirements, the model offers two operating modes: high and xhigh. In the flagship Agent mode (high), which emphasizes multi-step execution and tool invocation, it achieves a PinchBench score of 87.60, surpassing GPT-5.4 xHigh and Gemini-3.1-Pro high, with a ClawEval test score of 63.82. In the deep deliberation mode for mathematical reasoning and scientific research (xhigh), it scores 95.83 on AIME 26 and 88.27 on GPQA Diamond.

The official statement highlights that the computational requirements for text format conversion and mathematical competitions are vastly different, and this mechanism is designed to reduce token costs, enabling the model to serve as the default base for high-frequency scenarios such as tool orchestration, programming, and multi-round interactions. The model is now offering a free API trial on the OpenRouter platform in collaboration with Novita for one week (until May 15), and will soon open-source its weights.