Slow Responses Created a Bottleneck, and Cerebras Cut It
You use OpenAI's GPT-5.6 Sol for real work.
But response wait time gets in the way.
The more repeated requests you make, the more wait time becomes cost.
Cerebras solved this with Ultrafast mode.
According to the official announcement, it is the same GPT-5.6 Sol model.
It generates up to 750 tokens per second.
That is up to 14x faster than Standard.
750 Tokens/Sec: Speed Confirmed by a Benchmark
This figure is not marketing copy.
It was confirmed through official OpenAI and Cerebras materials.
It is a result on the Humanity's Last Exam benchmark.
It finished in 11 hours 11 minutes.
The comparison was Claude Fable5.
It took 78 hours 27 minutes.
That's about a 7x difference.
The model itself is unchanged.
Only the inference infrastructure changed.
It means you get the same answer faster.
| Mode | Throughput | Multiple |
|---|---|---|
| Standard | Baseline speed | 1x |
| Ultrafast | Up to 750 tokens/sec | Up to 14x |
12 Minutes Down to About 1? Not Quite
Cerebras also released its own demo.
It built a finance dashboard with the same prompt.
Ultrafast took 1 minute 50 seconds.
Standard took 12 minutes 20 seconds.
That's roughly a 7x faster result.
But this figure is Cerebras' own demonstration.
It is not an official OpenAI benchmark.
It's accurate to treat it as reference material only.
It is still useful for gauging prototype iteration speed.
Ultrafast Is Still Invite-Only
Ultrafast is still in a limited preview stage.
It is open only to some select customers through the OpenAI API.
General users can't use it right now.
Pricing has not been disclosed.
First, check whether you can get access.
Also check the date of the official announcement.
Before adopting it, look at the benchmark conditions too.