Overview
On August 13, 2026, OpenAI previewed Ultrafast, a new application programming interface (API) service tier that runs GPT‑5.6 Sol — the company's most capable large language model (LLM) — at up to 14 times the speed of its Standard processing tier. <cite index="4-3,4-4">The new tier runs GPT‑5.6 Sol up to 14× faster than Standard processing and is powered by Cerebras, generating up to 750 output tokens per second.</cite>
For reference, <cite index="6-2,6-4">Standard processing generates approximately 53 tokens per second.</cite>
Hardware Architecture
The speed differential is rooted in the underlying silicon. <cite index="1-3">Cerebras' Wafer-Scale Engine chips keep 44 GB of SRAM on-chip, avoiding the memory bandwidth bottlenecks typical graphics processing unit (GPU) setups face.</cite> <cite index="8-9">Traditional GPU-based systems must continuously shuttle massive model weights between external HBM and processing cores, while multi-GPU setups add cross-chip interconnect latency through PCIe or NVLink.</cite>
Ultrafast is not a new model. <cite index="9-4">It is a new way to serve an existing one</cite>, made possible by the compute architecture Cerebras brings to OpenAI's platform.
Tiered Inference Pricing Strategy
<cite index="3-3,3-4,3-5">OpenAI already monetizes inference speed in tiers. Through the API, it offers a "Fast Mode" that promises up to 2.5× speed with lower latency for GPT-5.6 Sol at roughly double the price. Ultrafast adds a third, faster, and likely pricier tier.</cite> <cite index="6-7">Ultrafast currently has no published pricing.</cite>
Underlying Partnership
Ultrafast is the first public product to emerge from a major infrastructure deal signed earlier in the year. <cite index="19-1">In January 2026, OpenAI signed a landmark $10+ billion agreement with Cerebras Systems to secure 750 megawatts of computing power through 2028.</cite> <cite index="22-12">"Cerebras adds a dedicated low-latency inference solution to our platform," Sachin Katti, who works on compute infrastructure at OpenAI, wrote in the blog.</cite>
Availability and Early Customers
<cite index="4-12,4-13">GPT‑5.6 Sol on Ultrafast mode is available in a limited preview to a select group of customers, with access to expand as capacity grows.</cite> <cite index="8-4">Early adopters include Jane Street, Basis, Rogo, and Podium.</cite> <cite index="14-9">OpenAI sees Ultrafast supporting live or near-production tasks including voice, customer support, commerce, developer agents, financial research, and security response.</cite>
Benchmark Performance
Cerebras published independent benchmark data alongside the announcement. <cite index="8-5">On the Humanity's Last Exam benchmark, Ultrafast completed all 2,500 questions in 11 hours and 11 minutes, compared with 78 hours and 27 minutes for Claude Fable 5.</cite> <cite index="2-1">On GDP-Val, a benchmark for economically valuable knowledge work tasks, Ultrafast delivered a 5.6× end-to-end speedup with no quality degradation.</cite>
Context: GPT-5.6 Family
<cite index="11-1,11-2">GPT-5.6 is an LLM developed by OpenAI and released on July 9, 2026, as a family of models in three variants ranked from least to most capable: Luna, Terra, and Sol.</cite> <cite index="16-2">GPT‑5.6 Sol sets a new standard for both intelligence and efficiency, achieving state-of-the-art results across coding, knowledge work, cybersecurity, and science.</cite> Ultrafast does not alter the model's weights or capabilities; it addresses only the serving layer.
OpenAI has not announced a general availability date or pricing for Ultrafast. Interested developers can register for updates through a waitlist form on OpenAI's website.