97.7 on Online Mine 2Web 83.2 on internal 68.6 on WebTail Bench Best across board and at 2.37 dollars per million token 90% Than GPT5.5 ! how they have trained this. They are using an undisclosed base model and using SFT to accelerate time to market, combined with asynchronous reinforcement learning, especially leveraging the GRPO algorithm. If you don't know, this is similar to how DeepMind historically has trained their AlphaGo Even though they are talking about 1-3 sec latency the ...