OpenAI Launches Cerebras-Supported Ultrafast Mode, Up to 14 Times Faster Than Standard Processing
OpenAI has released the Ultrafast service tier, running the GPT-5.6 Sol model, powered by Cerebras, with a generation speed of up to 750 output tokens per second.
This mode is up to 14 times faster than standard processing and is initially available in limited preview to selected customers via the OpenAI API, with plans to expand access as capacity increases.
Target scenarios include real-time voice, customer support, e-commerce, programming and design agents, financial research, and security response, all requiring low-latency enterprise workflows.
The Cerebras wafer-scale engine keeps model weights in on-chip SRAM, reducing data transport bottlenecks and achieving a combination of cutting-edge model intelligence and ultra-low latency.
Internally, OpenAI has used it to compress research cycles and incident log analysis, shortening tasks that previously took overnight to multiple iterations within a workday.
In market mechanisms, OpenAI embeds cutting-edge models directly into real-time products through dedicated high-speed computing, directing funds and demand towards low-latency inference infrastructure; beneficiaries include enterprise customers needing instant responses and Cerebras, while traditional GPU cluster-dependent high-latency inference service providers face pressure.
Source: Public Information
ABAB AI Insight
OpenAI has previously established a large-scale low-latency computing partnership with Cerebras, incorporating wafer-scale chips into the inference stack. This time, it directly deploys the strongest model, GPT-5.6 Sol, on it, continuing the path from general GPU clusters to dedicated high-speed hardware, consistent with the earlier strategy of separating training and inference.
On the capital path, OpenAI locks in Cerebras capacity through long-term computing agreements, productizing speed itself and charging enterprises through tiered services; the motivation is to seize the real-time interaction entry point, strategically using hardware differentiation to consolidate model advantages while reducing dependence on a single GPU supplier.
Similar cases can be seen with dedicated inference chips from Groq, SambaNova, and the low-latency instances launched by cloud vendors; the current AI infrastructure is in an expansion phase from "fast enough" to "real-time available," with speed becoming as important a competitive dimension as intelligence.
The structural judgment belongs to technological substitution: wafer-scale architecture replaces inter-chip data transport with on-chip memory, directly rewriting the upper limit of inference latency; the mechanism is that when model capabilities approach platform maturity, end-to-end response speed becomes a key variable determining whether products can enter real-time scenarios, thus redistributing control over computing power and application entry points.
ABAB News · Cognitive Laws
Once speed is fast enough, it becomes a new form of intelligence.
Dedicated hardware will ultimately consume the profits of general clusters.
Real-time response rights determine who defines the next generation of interaction.