OpenAI launches Ultrafast mode for GPT-5.6 Sol, delivering 750 tokens/second

OpenAI announced the preview launch of Ultrafast mode, a new tier of inference acceleration for its GPT-5.6 Sol model that delivers up to 750 output tokens per second. The capability is powered by Cerebras, the AI infrastructure company with which OpenAI signed a ten-billion-dollar partnership earlier in the year. The service will roll out first through the OpenAI API, restricted to select customers, with plans to expand access gradually as capacity permits. Interested companies can request early access via a signup form.

The speed breakthrough opens several classes of applications OpenAI believes were previously impractical at real-time scales. In incident response, engineers can have logs, code changes, and reports analyzed while an outage is still unfolding, enabling faster diagnosis and repair. OpenAI reports already using the model internally for this purpose. In finance, traders and risk teams could evaluate market signals and flag suspicious activity as conditions shift. Customer support could resolve complex multi-step inquiries without delay, and e-commerce platforms could answer product questions, check inventory, and personalize recommendations before a customer abandons their session. Research teams can convert overnight batch experiments into interactive work—testing an idea, reviewing results, adjusting approach, and rerunning without interrupting the workflow.

OpenAI's pricing strategy treats speed as a performance tier analogous to cloud infrastructure pricing. The company already offers Fast Mode for GPT-5.6 Sol at roughly double the standard price for up to 2.5x speed and lower latency. Ultrafast becomes a third, faster tier at an implied further premium. This model mirrors how AWS and other cloud providers charge more for the same computational resource delivered at higher performance, applying the same logic to AI inference. If inference speed becomes a bottleneck across industries—particularly in finance, customer support, and operations—this tiered model gives OpenAI direct revenue capture from the economic gains that faster inference unlocks.