IA · 15 August 2026 · 5 min read
OpenAI launches Ultrafast: GPT-5.6 Sol reaches 750 tokens per second with Cerebras chips
In brief: OpenAI has previewed 'Ultrafast', a high-speed mode for its flagship GPT-5.6 Sol model that delivers up to 750 output tokens per second—a 14-fold increase over standard processing speeds. Powered by Cerebras specialized wafer-scale silicon, the capability targets mission-critical enterprise workflows such as cybersecurity incident response, high-frequency financial analysis, and low-latency automated support.
by Team Mocchi's
The artificial intelligence race is no longer solely contested on parameter counts or reasoning depth; execution throughput has become the decisive battleground. OpenAI has announced the preview release of Ultrafast, a high-performance operating mode designed for its premier model, GPT-5.6 Sol. The feature aims to drastically reduce generation latency, achieving an output throughput of up to 750 tokens per second—a fourteenfold acceleration compared to standard processing speeds.
Historically, achieving near real-time interaction forced developers to make a hard technical trade-off: compromising on reasoning capabilities by relying on smaller, distilled models to maintain responsiveness. With Ultrafast, OpenAI seeks to eliminate this compromise, offering full frontier reasoning capabilities within latency windows suitable for demanding real-time workflows.
Under the hood: the Cerebras silicon partnership
Behind this throughput leap lies a dedicated hardware infrastructure. As reported by TechCrunch, the Ultrafast mode is powered through a collaboration between OpenAI and chipmaker Cerebras, leveraging the latter's wafer-scale processor architecture designed to eliminate memory bandwidth bottlenecks during large language model inference.
The launch directly answers competitive moves from rivals such as Anthropic, which previously introduced fast inference tiers for Claude without reaching the token rates announced in this release. Ultrafast is currently available in preview to a select group of enterprise clients and API developers, with broader commercial rollout planned over the coming months.
Mission-critical workloads and the enterprise push
The primary beneficiaries of this capability are corporate environments running time-sensitive operations. Generating 750 tokens per second enables the instant production of dense technical documentation, actionable incident remediation playbooks, and complex analytical reports within sub-second thresholds. Key deployment sectors include cybersecurity incident response, live financial market analysis, seamless conversational voice systems, and automated customer operations.
This infrastructure enhancement aligns with a broader enterprise push across the ecosystem. As detailed by TechCrunch, OpenAI has concurrently partnered with IBM to train and certify tens of thousands of corporate consultants on its toolchain, embedding GPT-5.6 directly into enterprise consulting platforms and signaling that AI competition has firmly shifted toward production-grade business execution.
Mocchi's take
For technology leaders and engineering teams, delivering over 700 tokens per second on premier frontier models marks an important architectural shift. System designs have long required convoluted two-tier setups—routing latency-critical interactions to smaller models while offloading deep reasoning to slower batch pipelines. High-throughput inference on top-tier models simplifies software stacks and makes conversational voice interfaces and autonomous real-time decision agents viable in mission-critical business environments.