OpenAI introduced a new API service tier called Ultrafast that runs its flagship GPT-5.6 Sol model as much as fourteen times faster than the standard tier, reaching roughly 750 output tokens per second while preserving the same underlying intelligence as the regular version of the model. The speed comes from a partnership with Cerebras, whose wafer-scale chip architecture keeps the entire model's weights resident in on-chip SRAM rather than shuttling them back and forth from external memory, which is normally the bottleneck that slows down large model inference on conventional GPU clusters. For teams building products where response latency directly affects usability, this matters a great deal: voice assistants, live coding copilots, real-time translation, and interactive agents all suffer when a model takes multiple seconds to produce a first token, let alone a full response. A model that can complete lengthy reasoning chains in a fraction of the usual time opens up interaction patterns that previously required smaller, less capable models as a latency compromise. The tradeoffs are real, though: Ultrafast is currently available only to a limited group of customers in preview, capacity is constrained by Cerebras hardware availability, and pricing details for general availability haven't been finalized. It also reflects a broader shift in the AI infrastructure market, where inference speed is becoming as competitive a battleground as raw model capability, with Cerebras and Groq both positioning specialized inference silicon against Nvidia's GPU dominance. For developers evaluating which provider to build on, the emergence of speed-optimized tiers means the classic tradeoff between model quality and response time is loosening, at least for well-funded customers with access to preview programs. Expect competitors to respond with their own high-throughput offerings in the coming months.