OpenAI previewed Ultrafast mode, a new API service tier that runs GPT-5.6 Sol at up to 14 times the speed of standard inference.

The service delivers up to 750 output tokens per second, powered by Cerebras chips optimized for high-throughput AI inference. OpenAI said the speed boost comes from specialized hardware designed for rapid token generation without sacrificing model quality.

GPT-5.6 Sol represents OpenAI's latest reasoning model, building on the company's previous GPT-o1 series. The model maintains the same capabilities as the standard version while dramatically reducing latency for real-time applications.

Ultrafast mode targets enterprise customers running high-volume inference workloads, particularly those building conversational AI products or processing large document batches. The company said early testing showed consistent performance across different prompt types and lengths.

The partnership with Cerebras marks OpenAI's first major infrastructure collaboration outside traditional GPU providers. Cerebras specializes in wafer-scale processors designed specifically for AI workloads, offering an alternative to Nvidia's dominant H100 chips.

Pricing details for Ultrafast mode were not disclosed in the preview announcement. OpenAI said the service will initially launch in limited beta for select enterprise customers before broader availability.

Why speed matters for AI deployment

Latency remains a critical bottleneck for AI applications, particularly in customer-facing scenarios where users expect near-instant responses. Traditional inference can take several seconds for complex reasoning tasks, limiting real-world deployment options.

The 14X speed improvement could enable new use cases like real-time code generation, instant document analysis, and responsive AI tutoring systems. Enterprise customers have consistently cited inference speed as a key factor in AI adoption decisions.

OpenAI plans to expand Ultrafast mode to other models in its lineup, starting with GPT-4o and Claude-competitive reasoning models. The company expects general availability within the next quarter, pending infrastructure scaling tests.