OpenAI’s Ultrafast is a new API service tier for GPT-5.6 Sol, powered by Cerebras Systems and advertised at up to 750 output tokens per second—or up to 14× the speed of Standard processing. The catch is important: Ultrafast is still a limited preview for selected OpenAI API customers, not a general ChatGPT feature, and the headline number describes output throughput rather than the total time every task will take.
Ultrafast is a service tier, not a new model
OpenAI introduced Ultrafast on August 13, 2026 as a way to serve GPT-5.6 Sol through a faster inference setup. GPT-5.6 Sol remains the model; Ultrafast is the API service tier and infrastructure choice around it.
That distinction matters for developers. This is not a new model with a new set of capabilities. It is an attempt to make an existing model respond quickly enough for workflows where waiting several seconds—or several minutes—changes the experience.
Cerebras Systems provides the inference infrastructure through its wafer-scale architecture. OpenAI says selected customers are using the preview for incident response, voice customer support, coding, commerce, financial research, security response and live experimentation.
| Dimension | GPT-5.6 Sol with Ultrafast | Practical takeaway |
| Product identity | An OpenAI API service tier for GPT-5.6 Sol | Ultrafast is not a separate model |
| Advertised output speed | Up to 750 output tokens per second | A maximum throughput figure, not a promise that every task finishes 14× faster |
| Standard comparison | Up to 14× Standard processing speed | The comparison is against Standard processing, under the conditions behind the claim |
| Access | Limited preview for selected OpenAI API customers | It is not a general ChatGPT rollout |
| Infrastructure | Cerebras Systems wafer-scale inference architecture | The hardware partnership is central to the service’s design |
| Public price and launch status | No dedicated public Ultrafast price or general-availability date announced | There is no ordinary public purchase path described yet |
What 750 output tokens per second does—and does not—tell you
Tokens are chunks of text processed by a language model. Output tokens per second measures how quickly the model generates its answer after generation begins. It is useful, especially for long responses and interactive applications, but it is not the same thing as end-to-end latency.
A real request also includes network travel, request queuing, input processing, tool calls, model reasoning, safety checks and the time needed to produce the first token. Under load, the experience can depend on concurrency and tail latency—the slower responses near the end of a distribution, often described with p95 latency.
So the careful reading is this: OpenAI and Cerebras advertise a maximum of up to 750 output tokens per second and up to 14× Standard processing speed. Those figures do not establish a sustained rate for every request, a p95 response time, a concurrency guarantee or a 14× reduction in total task time.
That is also why output speed should not be confused with cost per completed task. A system that writes faster may still spend time waiting for tools, handling a large context or completing several steps in an agent workflow. Fast typing is excellent; it is not the same as finishing the whole job instantly. Sadly, computers have not yet abolished paperwork.
Why Cerebras is part of the story
Cerebras says its wafer-scale architecture keeps model weights on-chip and provides 44 GB of SRAM per wafer-scale chip for the cited system. SRAM is fast memory located close to the processing hardware. In this design, the company presents that memory architecture as part of the explanation for Ultrafast’s low-latency inference.
The relevant product family is Cerebras’s Wafer-Scale Engine, which places a very large amount of computing hardware on a single wafer-sized platform rather than spreading the workload across conventional accelerator cards in the usual way. For an API customer, the important point is less the shape of the chip than the intended result: the model-serving infrastructure is optimized around rapid inference.
Cerebras has also published its own benchmark results for Ultrafast, including a reported 5.6× end-to-end speedup on GDP-Val without reported quality degradation. Those figures are company benchmarks, so they provide context for the vendor’s performance case rather than an independent, matched test of every production workload.
The workflows where lower latency could matter
The strongest case for Ultrafast is not simply “the chatbot answers faster.” It is the possibility of keeping a person or an automated workflow in a tighter loop with the model.
OpenAI’s demonstration shows an interface investigating a security incident with logs, metrics and file operations, followed by a code-refactoring workflow. The interface displays a model indicator for “5.6 Sol Ultra” and tracks the work as it proceeds. The demonstration is useful because it shows the interaction pattern: the model is repeatedly asked to inspect information, reason about it and produce another step without forcing the user to wait through long pauses.
In that demonstration, AJ Trbojevic says a security workflow that previously took one to two hours to collect, sort, normalize and contextualize data took roughly 10–15 minutes with Ultrafast. That is an official demonstration result, not an independent production benchmark, but it illustrates why lower latency matters in incident response.
The same logic applies to the other use cases OpenAI lists. Faster turns can make voice support feel more conversational, allow developers to iterate more quickly while coding, support synchronous commerce experiences and make financial research interfaces feel more interactive. They may also help live experiments where a team needs to try several prompts or approaches in succession.
That does not establish suitability for high-frequency trading, exchange connectivity, deterministic execution or regulatory compliance. The announcement supports time-sensitive financial research as a use case; it does not turn Ultrafast into a trading system.
Access, price and the unanswered production questions
For readers in the United States and elsewhere, the practical status is straightforward: Ultrafast is a limited preview for selected OpenAI API customers. The announcement does not provide a general-availability date, and it does not describe Ultrafast as a feature available to ordinary ChatGPT users.
The launch information does not provide a dedicated public Ultrafast price. Standard API pricing should not be treated as Ultrafast pricing. That leaves an important part of the business case unresolved for developers: whether the speed improvement will justify the eventual cost for a particular workload.
Before adopting a service like this, an engineering team would want to measure more than the fastest generated output. Useful production questions include:
- How quickly does the first token arrive under realistic load?
- What does p95 latency look like when many customers use the service at once?
- How does performance change with long inputs, tool calls and multi-step agents?
- What is the cost per completed workflow rather than per generated token?
- Does the model preserve the quality and behavior expected from GPT-5.6 Sol across the team’s own tasks?
Those are the measurements that turn an impressive throughput claim into an infrastructure decision.
The bottom line
OpenAI’s Ultrafast is a meaningful change in how GPT-5.6 Sol can be served: a Cerebras-powered API tier aimed at making demanding, interactive workflows feel much less sluggish. The advertised ceiling—up to 750 output tokens per second and up to 14× Standard processing—is attention-grabbing, but it is not a universal end-to-end speed guarantee.
For now, Ultrafast is best understood as a limited preview for selected API customers. Its real value will depend on access, eventual pricing, loaded latency and the time required to complete useful work—not just how quickly text appears on the screen.
