OpenAI previewed a new API service tier called Ultrafast on August 13, promising speeds of up to 750 output tokens per second for its GPT-5.6 Sol model. That rate is 14 times faster than standard inference on the same model.
The tier is powered by chips from Cerebras Systems rather than the Nvidia GPUs that underpin most of OpenAI’s existing infrastructure.
Key Takeaways
OpenAI previewed a service tier called Ultrafast on August 13, offering up to 750 output tokens per second for GPT-5.6 Sol
The 750-tokens-per-second rate is 14 times faster than standard inference on the same model
The Ultrafast tier runs on chips from Cerebras Systems rather than the Nvidia GPUs used in most of OpenAI’s infrastructure
A 10,000-token response takes about three minutes and twenty seconds at 50 tokens per second versus thirteen seconds at 750
The preview signals that OpenAI is willing to route production workloads through third-party silicon when speed requirements exceed what its own data centers can deliver.
OpenAI Ultrafast API Breaks From Nvidia-Only Infrastructure
OpenAI’s own lab blog post published the details. The company described the OpenAI Ultrafast API as a preview tier available to API developers, with GPT-5.6 Sol as the initial model.
The 750-tokens-per-second figure represents peak throughput under favorable conditions. Typical sustained output will vary by prompt complexity and concurrent load.
OpenAI did not disclose pricing for the OpenAI Ultrafast API in the initial announcement.
The significance of the Cerebras partnership lies in the hardware architecture. Conventional AI inference runs on graphics processing units arrayed in clusters, where the bottleneck is the time required to move model weights between high-bandwidth memory and compute cores.
Cerebras’s wafer-scale design eliminates the inter-chip communication overhead that limits cluster-based inference, allowing weights to live in on-chip memory and be accessed in a single clock cycle rather than shuffled across interconnects. The practical result is dramatically lower latency per token, which is why Cerebras hardware consistently tops third-party speed benchmarks for large language model inference even though the company’s training capabilities lag Nvidia’s.
For OpenAI, routing GPT-5.6 Sol through Cerebras infrastructure lets it offer a tier that no GPU cluster could match on raw token throughput, without requiring OpenAI to redesign its own data center fleet.
Why 750 Tokens Per Second Matters For Real Applications
The difference between 50 tokens per second and 750 tokens per second is not merely cosmetic. At 50 tokens per second, a 10,000-token response takes about three minutes and twenty seconds to complete.
At 750 tokens per second, the same response arrives in thirteen seconds. For agentic workflows where an AI model iterates through dozens of reasoning steps before returning an answer, that compression is the difference between a product that feels interactive and one that feels like batch processing.
OpenAI’s own research published in August showed that enterprises deploying agentic AI see the widest productivity gains in tasks requiring rapid sequential inference, such as code review loops, multi-step document processing, and real-time customer support.
Faster inference makes those loops economically viable at scale because fewer parallel compute instances are needed to serve the same number of users.
The 14X figure also carries competitive weight. Anthropic, Google DeepMind, and independent inference providers including Groq and Together AI have each marketed inference speed as a differentiation axis. The OpenAI Ultrafast API’s entry into ultra-high-throughput inference via Cerebras hardware raises the ceiling for what enterprise customers can demand, and forces competitors to respond on the same dimension rather than routing the conversation back to model quality alone.
From OpenAI’s GPU Dependence To A Multi-Silicon Strategy
OpenAI’s relationship with Nvidia has historically been close to exclusive on the inference side.
The company’s deal structures with Microsoft Azure, which supplies its primary compute, are built around Nvidia A100 and H100 clusters. Routing production traffic through Cerebras is a meaningful departure.
The move parallels what happened in the memory market earlier this year, when hyperscalers began qualifying second-source suppliers to reduce dependence on a single chipmaker.
OpenAI is not abandoning Nvidia. The vast majority of its inference will remain GPU-based.
The OpenAI Ultrafast API is better understood as a targeted product for latency-sensitive workloads rather than a wholesale infrastructure shift.
Cerebras has sought a public listing and disclosed revenue figures showing rapid growth, though hardware revenue has been volatile quarter to quarter. A high-profile production partnership with OpenAI provides the kind of reference customer validation that enterprise semiconductor buyers require before making large procurement commitments.
The arrangement benefits both parties in ways that go beyond the immediate API tier.
What The OpenAI Ultrafast API Signals For Pricing Architecture
OpenAI currently structures its API around two primary axes: model capability and context length. The OpenAI Ultrafast API adds a third axis, throughput, and implies that OpenAI will charge a premium for it.
That is a structurally important shift because it allows OpenAI to extract higher revenue per token from latency-sensitive enterprise customers without lowering prices for standard-speed users.
The parallel is cloud computing’s reserved versus on-demand instance model. Standard inference is on-demand, priced per token at a rate that reflects average cluster utilization.
The OpenAI Ultrafast API implies a dedicated or prioritized allocation of Cerebras capacity, which carries a higher unit cost that enterprise customers with real-time requirements will pay willingly.
If the preview proves stable, expect OpenAI to formalize the OpenAI Ultrafast API as a permanent tier and extend it to additional models beyond GPT-5.6 Sol. The initial restriction to one model is consistent with how OpenAI has historically launched infrastructure features: preview on a single surface, gather production data, then generalize.
The 14X speed claim is the headline number, but the structural shift toward multi-silicon inference architecture is the development with longer consequences.
Read Next: SK Hynix Bets $720 Billion on AI Memory, Aiming to Triple Capacity by 2034