a daily news desk
Tools

OpenAI, Cerebras preview 'Ultrafast' tier: GPT-5.6 Sol at 750 tokens/sec

The limited-preview API service runs OpenAI's flagship reasoning model on Cerebras wafer-scale chips at a vendor-claimed 14x standard speed.

By Iris Holloway · Tools & frameworks ·

OpenAI and Cerebras on August 13, 2026 unveiled Ultrafast, a limited-preview API tier that runs GPT-5.6 Sol at up to 750 output tokens per second, a vendor-claimed 14x the throughput of OpenAI’s Standard tier. No new model, no new context window, no new intelligence: the same GPT-5.6 Sol that reached general availability on July 9, moved onto Cerebras Wafer-Scale Engine silicon and its 44 GB of on-chip SRAM.

Pricing hasn’t been disclosed. Neither has a general-availability date. Standard GPT-5.6 Sol still runs at $5 per million input tokens and $30 per million output tokens; Ultrafast customers get latency, and OpenAI gets telemetry.

“We’re starting with a small group of customers to learn where that speed creates meaningful value,” said Sachin Katti, VP Compute Strategy & GPT-Infra at OpenAI. The named verticals are coding, commerce, financial research, support, and voice AI, the categories where sub-second reasoning is a product feature rather than a demo.

The benchmark theater is Cerebras’s. On Humanity’s Last Exam, a 2,500-question set, GPT-5.6 Sol on Ultrafast finished in just over 11 hours; Claude Fable 5, at comparable accuracy, took more than three days. On GDP-Val, Cerebras claims a 5.6x end-to-end speedup with no loss in quality. Every headline number here comes from one of the two companies shipping the product. No independent third party has benchmarked it.

What Ultrafast really previews is the commercial output of the January 2026 infrastructure deal between the two companies: 750 megawatts of Cerebras capacity committed through 2028, 600 MW already contracted per the August 12 Q2 earnings, manufacturing scaling more than 10x during 2026. Techtimes puts the underlying arrangement at over $10 billion, including a $1 billion OpenAI loan secured by warrants on Cerebras shares.

The customer, in other words, is also the creditor, and holds paper on the supplier’s equity. Ultrafast is what that balance sheet looks like when it ships an SKU.

Sources