OpenAI has opened a preview of Ultrafast, a new API service tier that runs its most capable model, GPT-5.6 Sol, up to 14 times faster than standard inference. The company announced the launch Thursday, saying the mode delivers up to 750 output tokens per second.
The speed gains come from a partnership with chipmaker Cerebras. Ultrafast runs on Cerebras Wafer-Scale Engine hardware, which keeps 44 GB of SRAM on-chip to sidestep the memory-bandwidth bottlenecks that limit typical GPU clusters.
What's new
- Model: GPT-5.6 Sol (OpenAI's current flagship reasoning model)
- Speed claim: Up to 14× standard throughput; up to 750 output tokens/second
- Hardware: Cerebras Wafer-Scale Engine (WSE) chips
- Availability: Limited preview for a small group of customers; access will expand as capacity grows
- Target workloads: Incident response, customer support, financial market analysis, e-commerce
OpenAI framed the release as a step past the traditional trade-off where real-time latency required smaller or specialized models. "Ultrafast points to progress in a new direction: more useful work per second," the company said in its blog post.
The preview is accessible through the OpenAI API. Pricing for the Ultrafast tier has not been disclosed publicly.
Why it matters
Inference speed has become a primary differentiator for enterprise AI adoption. Latency-sensitive workflows — real-time trading analytics, live customer-support agents, automated incident triage — have often forced teams to downsize models or accept lower reasoning quality. Ultrafast aims to keep full GPT-5.6 Sol capability while meeting sub-second response requirements.
The Cerebras partnership also signals a shift in inference infrastructure. By moving the flagship model onto wafer-scale silicon, OpenAI gains a hardware path that does not rely solely on NVIDIA GPU supply chains, a consideration that matters to enterprise buyers evaluating vendor lock-in and capacity planning.
Our take
The 14× figure is impressive, but the preview's limited seat count means most developers cannot yet verify sustained throughput, tail latency, or cost per million tokens under real workloads. Until pricing and broader access arrive, Ultrafast remains a capability signal rather than a production option for the majority of teams.
