GPT-5.6 Sol Hits 750 Tokens a Second
The loudest number of the week attached to the thing you can least act on. Both halves matter.
Founder, Automation Squad ·
The short answer
OpenAI previewed Ultrafast on August 13, 2026 — a service tier running GPT-5.6 Sol on Cerebras hardware at up to 750 output tokens per second, described as up to 14× the speed of standard processing with unchanged model intelligence. It is a limited preview for selected customers. No price and no general availability date have been published.
The numbers, and the honest read on what they're worth
Every figure here is vendor-published. The most useful column is the last one, which is empty on purpose.
Work out whether latency is actually your constraint
For most workloads it is not — quality or cost is. Latency becomes the binding constraint in three places: voice, anything a human waits on live, and agent loops where each step blocks the next.
Measure your current tokens-per-second before you covet a bigger one
Time one real request end to end and divide by output tokens. Without your own baseline, 750 is a number with nothing to compare against.
Count the sequential steps in your slowest agent loop
Speed multiplies through sequential steps and does nothing for parallel ones. A ten-step chain is where a 14× improvement is transformative; a single call where a human then reads the output for two minutes is where it is irrelevant.
Do not rebuild anything around this yet
Limited preview, no price, no GA date. Any architecture that assumes access is an architecture assuming a commercial term nobody has published.
| Measure | Figure |
|---|---|
| Output speed | Up to 750 tokens/sec |
| Versus standard tier | Up to 14× faster |
| Model intelligence | Unchanged — same GPT-5.6 Sol |
| Humanity's Last Exam, 2,500 questions | 11h 11m (Claude Fable 5: 78h 27m) |
| GDP-Val | 5.6× end-to-end speedup, no quality degradation |
| Why it's fast | Cerebras keeps weights on-chip in 44 GB SRAM per wafer |
| Named early testers | Jane Street, Podium, Basis, Rogo |
| Price | Not disclosed |
| General availability | Not announced |
Take it with you
OPENAI ULTRAFAST — the numbers and the catch
Previewed Aug 13, 2026
Source: https://automationsquad.com/news/openai-ultrafast-cerebras/
THE NUMBERS (all vendor-published)
Output speed .................. up to 750 tokens/sec
vs standard tier .............. up to 14x
Model intelligence ............ UNCHANGED (same GPT-5.6 Sol)
Humanity's Last Exam .......... 2,500 questions in 11h 11m
(Claude Fable 5: 78h 27m)
GDP-Val ....................... 5.6x end-to-end, no quality loss
Mechanism ..................... Cerebras wafer-scale, weights on-chip
in 44 GB SRAM per chip
Early testers ................. Jane Street, Podium, Basis, Rogo
THE CATCH
Price ......................... NOT DISCLOSED
General availability .......... NOT ANNOUNCED
Access ........................ limited preview, selected customers
BEFORE YOU CARE
[ ] Is latency actually my constraint, or is it cost/quality?
[ ] What is my current tokens/sec on a real request? ______
[ ] How many SEQUENTIAL steps in my slowest agent loop? ______
(speed multiplies through sequential steps, not parallel ones)
[ ] Am I about to design around something with no published price?OpenAI previewed a new API service tier called Ultrafast on August 13, 2026, running GPT-5.6 Sol on Cerebras hardware. Cerebras published its own account of the engineering the same day.
The facts: output runs at up to 750 tokens per second, which both companies describe as up to 14× the speed of the standard tier, with model intelligence unchanged. Cerebras attributes the gain to its Wafer-Scale Engine holding weights on-chip in 44 GB of SRAM per wafer-sized chip, removing the memory-bandwidth bottleneck. Two benchmarks were published: all 2,500 questions of Humanity's Last Exam completed in 11 hours 11 minutes against Claude Fable 5's 78 hours 27 minutes, and a 5.6× end-to-end speedup on GDP-Val with no quality degradation. Availability is a limited preview to a select group of customers, with named early testers including Jane Street, Podium, Basis and Rogo. No pricing has been disclosed and no general availability date announced.
Automation Squad's take: this was the loudest number in its week and the least actionable, and both facts deserve equal billing. A 14× latency improvement with no quality cost is genuinely significant for a narrow set of workloads — voice, anything a person waits on in real time, and agent loops where each step blocks the next, because that is where speed compounds through the chain rather than shaving a second off one call. For everything else it changes nothing you can act on. And the absent number is the important one: a preview with no published price is not a product decision yet, it is a signal about where inference is heading. Treat it as weather, not as a tool.
Run this now: time one real request end to end and divide by its output tokens, so you have your own tokens-per-second baseline. Then count the sequential steps in your slowest agent workflow. If the answer is one or two, latency is not your problem and you can safely ignore this until it has a price. If the answer is eight, you have just found the place where a change like this would eventually matter — and you now know what to test the day it becomes buyable.
Questions people are asking
- How fast is OpenAI's Ultrafast mode?
- Up to 750 output tokens per second, which OpenAI and Cerebras both describe as up to 14× the speed of standard processing. Model intelligence, context window and output quality are unchanged — only latency differs.
- Can I use Ultrafast today?
- Almost certainly not. It is a limited preview available to a selected group of customers, with access expanding as capacity grows. Named early testers include Jane Street, Podium, Basis and Rogo.
- What does Ultrafast cost?
- No pricing has been published, and no general availability date has been announced. That is the single most important fact about it for planning purposes.
- Why is Cerebras hardware faster for this?
- Cerebras attributes it to its Wafer-Scale Engine keeping model weights on-chip in 44 GB of SRAM per wafer-sized chip, which removes the memory-bandwidth bottleneck that limits conventional inference hardware.
Sources
Further reading
- Cerebras — Accelerating GPT-5.6 Sol Ultrafast with OpenAI — The engineering detail, including why wafer-scale SRAM removes the bandwidth bottleneck.
Last checked August 13, 2026 against the primary sources above, by Automation Squad Research. Spot an error? [email protected].