On August 13, 2026, OpenAI began a limited preview of "Ultrafast," a new service tier for running GPT-5.6 Sol at high speed. It launches first through the OpenAI API, promising up to 14 times the speed of Standard processing and up to 750 output tokens per second. Previously, use cases demanding maximum speed often meant choosing smaller or specialized models. Ultrafast is an attempt to bring Sol—the model OpenAI considers its most intelligent—into the timeframes needed for incident response and voice conversations, without sacrificing that intelligence.
However, availability is currently limited to a preview for select customers. Neither pricing nor a general availability date has been announced. How much the stated figure of up to 750 tokens/second will actually reduce end-to-end latency for real workloads needs to be measured including the time a model spends waiting on external tools and processing input.
Up to 750 Tokens/Second Breaks Down the Speed-vs-Intelligence Tradeoff
::tweet[https://x.com/OpenAI/status/2087947721936359705?s=20]
What Ultrafast changes isn't the model itself, but the speed tier running that same GPT-5.6 Sol. OpenAI describes it as up to 14 times faster than Standard processing, with generation speeds reaching up to 750 output tokens per second. It's rolling out first through the OpenAI API. Early customers in coding, commerce, financial research, and customer support are testing it.
The difference from the existing Fast mode is substantial. On July 30, 2026, OpenAI renamed Priority processing to Fast mode, which speeds up GPT-5.6 Sol by up to 2.5 times compared to Standard. Fast mode costs twice the Standard price and can be activated by specifying fast or priority in the service_tier field of an API request. Ultrafast's pricing and how to specify it, meanwhile, have not yet been disclosed.
| Speed Tier | Announced Speed for GPT-5.6 Sol | Availability | Announced Pricing Terms |
|---|---|---|---|
| Standard | Baseline for comparison | General availability | $5 per million input tokens, $30 per million output tokens |
| Fast | Up to 2.5x faster than Standard | Available via API | 2x Standard pricing |
| Ultrafast | Up to 14x faster than Standard, up to 750 output tokens/sec | Limited preview for select customers | Not announced |
This comparison shows that Ultrafast isn't so much a replacement for existing fast options as an access-restricted, low-latency path layered on top of them. Under Fast mode, a sudden surge in traffic can cause some requests to be downgraded to Standard. The same capacity constraints appear to govern Ultrafast, as OpenAI states it will expand access "as capacity grows."
Cerebras's 750MW Plan Reaches a Flagship Model
In January 2026, OpenAI and Cerebras announced a partnership to integrate 750MW of low-latency AI compute into OpenAI's infrastructure. The compute capacity is planned to come online in multiple phases through 2028, as part of OpenAI's strategy of matching workloads to the right compute systems. Cerebras consolidates compute, memory, and bandwidth onto a single massive chip, reducing the data-movement bottlenecks that occur during inference on conventional hardware.
The first product to use this was GPT-5.3-Codex-Spark in February 2026—a smaller version of GPT-5.3-Codex that generates over 1,000 tokens per second on Cerebras's Wafer Scale Engine 3. It launched as a text-only research preview with a 128k context window, narrowly targeted at fast coding tasks.
With Ultrafast, that low-latency path has now extended to GPT-5.6 Sol. While its generation speed falls short of Codex-Spark's 1,000+ tokens per second, what matters for comparison is the model's scale and role. This marks a progression from running a small, specialized model quickly to delivering responses from a flagship model handling complex, specialized work at up to 750 tokens per second. The 750MW compute plan announced in January is beginning to take shape as an API service built on Sol.
Speed Matters More for Iteration Count Than Response Display
Among the use cases OpenAI highlighted, incident response makes the value of speed most tangible. During a system outage, the model reads logs, traces, and recent code changes, then updates its hypothesis about the root cause and the next things to check. Engineers still own the judgment calls and deployments, but the interval between spotting a signal and testing a hypothesis can be compressed.
OpenAI is pursuing the same effect in research. Work that used to mean running a batch of experiments overnight and checking results the next morning is now being tested as something that can be cycled through repeatedly within a single workday. Financial research, commerce, and voice-based customer support handling complex inquiries all depend on whether the system can move to the next step before circumstances change. Speed matters less for displaying answers faster and more for increasing how many attempts a person or piece of software can make in a day.
There's a clear boundary here, too. Output token speed is a measure that applies only to the interval when the model is generating a response. OpenAI describes Codex-Spark's total task time as the sum of output generation plus input prefill, tool execution, and network time. If an external database's response or a payment process is slow, choosing Ultrafast won't make the overall task 14 times faster. Adoption benefits should be measured not by token speed but by the time from when a user initiates an action to when they receive a result.
Verify Capacity and Real-World Measurements Before Pricing
As of August 13, 2026, Ultrafast remains limited to a small number of customers. OpenAI says it will expand access as capacity grows, but hasn't disclosed a general availability date or supported regions. Rate limits and service-level details also remain unknown. The published figures—"up to 14x" and "up to 750 output tokens/second"—leave open questions about average speed and how performance changes with longer inputs.
There are three numbers enterprises should verify: time to first token, total task completion time including tool processing, and sustained output speed under load. Layering Ultrafast's per-unit pricing on top of these figures will help distinguish workloads that are adequately served by Fast mode's up-to-2.5x speedup from those worth paying a premium for even greater speed.
OpenAI cited evaluations from early customers Jane Street and Podium, along with comments from Basis and Rogo. However, none of these constitute controlled comparative testing under uniform measurement conditions. Whether Ultrafast becomes a viable option for production systems will be determined once pricing is disclosed during the access expansion and companies measure latency using their actual prompts and tool configurations. Until the 750MW of compute comes fully online in phases through 2028, the pace of capacity growth deserves just as much attention as the speed figures themselves.