OpenAI and Cerebras Push GPT-5.6 Sol to 750 Tokens per Second, Collapsing the Speed-Intelligence Tradeoff

1. OpenAI and Cerebras Push GPT-5.6 Sol to 750 Tokens per Second, Collapsing the Speed-Intelligence Tradeoff

OpenAI previewed Ultrafast on August 13, 2026: a new API service tier running GPT-5.6 Sol at up to 750 output tokens per second, which the company says is 14 times faster than its Standard processing tier. The tier is powered by Cerebras hardware and is currently in limited preview with an initial cohort of customers across coding, commerce, financial research, and voice applications. Access is gated; companies outside the preview can sign up for a waitlist. Early customers include Jane Street, Podium, and Basis.

The strategic move here is not speed for its own sake. Until now, real-time latency in production meant accepting a smaller or more specialized model, a concession that kept frontier intelligence out of the most time-sensitive workflows. Ultrafast breaks that constraint directly. Competitors like Anthropic, Google DeepMind, and Meta all face the same architectural tension: serving a frontier model fast enough for live voice, incident response, or fraud detection without degrading to a smaller distilled variant. OpenAI's partnership with Cerebras gives it a differentiated inference path that the hyperscaler-hosted competition cannot easily replicate on standard GPU clusters. For enterprise buyers evaluating agent infrastructure, this shifts the comparison frame from "which model is smartest" to "which provider can run the smartest model in real time."

The Cerebras pairing is worth watching as a template. OpenAI is effectively treating specialized silicon as a product tier rather than an internal cost lever, which signals that inference infrastructure is becoming a customer-facing competitive dimension, not just an ops concern. If Ultrafast moves from preview to general availability with stable pricing, it sets a new baseline expectation for what a frontier API should deliver, and puts pressure on every provider still asking customers to choose between speed and capability.

Source: Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed