GPT-5.6 Sol hits 750 tokens per second as OpenAI deploys Cerebras chips

Speed beyond standard hardware

OpenAI has started testing a new API tier named Ultrafast, promising to accelerate the GPT-5.6 Sol model up to fourteen times its current baseline. According to the announcement, the tier generates up to 750 output tokens per second, powered by a shift to Cerebras hardware. The primary announcement page was blocked during verification, so I am carefully relying on metadata and the initial signal rather than unverified technical documentation.

Who actually needs this throughput

When an LLM streams 750 tokens per second, it far exceeds human reading speed. This throughput is no longer built for chat windows. It targets real-time speech synthesis, seamless multimodal interactions, and agents that need to evaluate complex decision trees in a fraction of a second. It shifts the economics and technical boundaries for applications that previously failed on latency.

The catch of specialized compute

Moving from generic GPUs to Cerebras architecture delivers a massive leap in throughput, but it raises immediate questions about capacity constraints and pricing. A specialized tier typically signals that this performance level will not be broadly or cheaply available. The Ultrafast tier is likely to remain a premium privilege for very specific enterprise workloads.

Waiting for production stress

The crucial metric to watch is whether the 750 tokens per second benchmark holds up under real production loads and parallel request spikes. Once developers start hitting the API during peak hours, we will see if the chips can maintain their promised stability.

Lilith's verdict

The shift to Cerebras chips proves the race is no longer about having a smart model, but about brute-forcing the output before the user even takes a breath.