OpenAI just made its most powerful model a lot faster. The company announced a limited preview of “Ultrafast mode” for GPT-5.6 Sol on August 13, delivering processing speeds up to 14 times faster than the standard version of the same model. The target audience: enterprise customers who need AI that can keep up with real-time workflows like voice applications, financial research, and security response.
The key number is 750 output tokens per second. For context, that’s roughly the equivalent of generating an entire page of text in about one second, fast enough that the bottleneck in most applications shifts from “waiting for the AI” to “figuring out what to do with the answer.”
What Cerebras brings to the table
The speed gains aren’t coming from software tricks alone. OpenAI partnered with Cerebras, the AI chip company known for building wafer-scale processors that dwarf conventional GPUs, to power the Ultrafast tier. Cerebras hardware enables GPT-5.6 Sol to run 11 times faster than Fable 5, OpenAI’s previous-generation model, and 5 times faster than Opus 4.8 running on Fast mode.
Why 14x speed matters for enterprise AI
OpenAI is explicitly positioning Ultrafast mode for mission-critical applications. Real-time voice is perhaps the most obvious beneficiary. Voice assistants powered by large language models have historically struggled with the awkward pause between a user finishing a sentence and the AI responding. At 14x standard speed, that gap shrinks to something approaching natural conversation.
Financial research is another target. A general-purpose model running at Ultrafast speeds could potentially replace or augment bespoke NLP systems used by hedge funds and trading desks to parse earnings calls, regulatory filings, and market data in real time.
Security response rounds out the marquee use cases. When a breach is detected, the speed at which an AI system can analyze logs, identify attack vectors, and recommend containment steps directly correlates with damage mitigation.
The competitive landscape is heating up
GPT-5.6 Sol itself launched in July 2026, representing what OpenAI described as a leap forward in coding, cybersecurity, and scientific research capabilities. The Ultrafast mode announcement comes barely a month later.
The initial rollout is deliberately constrained. Only select customers will get access through the OpenAI API, with broader availability planned once capacity scales up.