OpenAI’s new Ultrafast tier shows that frontier AI is competing not only on intelligence, but also on speed. Powered by Cerebras, GPT-5.6 Sol is claimed to deliver up to 750 tokens a second — nearly 14 times faster than standard processing. It is currently an invite-only API preview aimed at time-sensitive business tasks.
Such speed could help cybersecurity teams, engineers, financial analysts and customer-support systems respond almost instantly. But faster answers are not necessarily better answers. Accuracy, reliability, privacy and human oversight will matter more than impressive demonstrations or benchmark figures.
The real question is not how many tokens a model produces per second, but how much useful work it delivers for the cost of computing, electricity and human review. Most companies may need cheaper models for routine tasks, faster systems for interactive work and the most powerful models only for complex or high-risk decisions. Research shows that accuracy and runtime often involve a trade-off.
OpenAI’s partnership with Cerebras and its planned 750 megawatts of low-latency compute indicate serious infrastructure ambitions, but not universal business need. Enterprises should demand independent production results, compare costs and retain human checks wherever errors can affect money, safety, rights or reputation.
SPEED CAN AMPLIFY GOOD JUDGMENT—BUT IT CAN ALSO AMPLIFY BAD DECISIONS.
Sanjay Sahay
Have a nice evening.