Contents
OpenAI previewed a new API tier this week called Ultrafast, and the headline number is the kind that changes what you’re willing to build. Running on Cerebras wafer-scale silicon instead of standard GPU infrastructure, GPT-5.6 Sol on Ultrafast generates up to 750 output tokens per second, roughly 14 times the ~53 tokens per second the standard tier delivers. That is the difference between a model that streams an answer while you wait and a model that finishes before you have registered it started.
Access is a limited preview for now, rolling out to a select group of customers with capacity expanding over time. But the architecture behind it, and the workloads OpenAI is targeting first, tell you where this is headed for anyone building on top of frontier models.
Three things worth understanding
The speedup comes from keeping weights on-chip, not from a smaller model. Cerebras packs 44 GB of SRAM directly onto its wafer-scale chip, large enough to hold GPT-5.6 Sol’s weights without shuttling them across a memory bus for every forward pass. Standard GPU inference is bottlenecked by memory bandwidth long before it is bottlenecked by compute; Cerebras’s bet is that removing that bottleneck lets the same model run an order of magnitude faster without any quality tradeoff. Unlike faster-but-smaller distilled variants, this is the full GPT-5.6 Sol, just served on different hardware.
Pricing for the speed premium is still unknown. Standard GPT-5.6 Sol lists at $5 per million input tokens and $30 per million output tokens. OpenAI has not published what Ultrafast costs on top of that, and given Cerebras hardware is scarcer and more expensive to operate than commodity GPU fleets, expect a real premium once general availability lands. Anyone planning to build on this tier should treat the current preview pricing (free or bundled, per early reports) as a promotional rate, not a stable number to architect around.
The target workloads are latency-critical, not throughput-critical. OpenAI is testing Ultrafast specifically on incident response and reliability tooling, plus financial research and security workflows, cases where a multi-second gap between “something is wrong” and “here is the analysis” has real cost. That is a different sales pitch than most inference speedups, which usually chase cheaper batch throughput. Ultrafast is explicitly about collapsing the gap between “smart enough to trust” and “fast enough to sit inside a live decision,” which is the tradeoff agent builders have been stuck making since GPT-4 shipped.
Try it: standard vs Ultrafast latency
The number that matters when you’re deciding whether a workload belongs on a fast tier is simple: how long does the model actually take to finish. Enter an output token count below and compare standard-tier latency against the Ultrafast ceiling.
About the demo
This uses the published throughput figures (53 tokens/sec standard, 750 tokens/sec Ultrafast) to estimate generation time. It does not call a live OpenAI endpoint.Enter a token count and click Estimate.
Where this leaves API consumers
My take is that Ultrafast is less about any single workload and more about a threshold effect: below a certain latency, entire categories of product become viable that were not before, live agent narration, synchronous incident triage, in-call financial analysis. Above that threshold, a smarter but slower model is just a worse version of asking a human to wait. If Cerebras and OpenAI can hold 750 tokens per second at general availability pricing that is not punitive, expect every other lab racing frontier benchmarks to also start racing latency, and expect “which tier for which workload” to become as routine an engineering decision as choosing a database index. The practical move today, if you are nowhere near preview access, is to start cataloguing which of your own agent calls are actually latency-bound versus quality-bound, so you already know which ones to move the moment a fast tier becomes generally available.