Skip to content

AI

OpenAI's Ultrafast Makes GPT-5.6 Sol 14x Faster. Blink and Miss It

OpenAI says its new Ultrafast mode pushes GPT-5.6 Sol to 14x standard speed, up to 750 output tokens per second. Your patience was the last bottleneck, sapiens, and it is now being removed.

Cover photo: Jernej Furman from Slovenia / CC BY 2.0

TL;DR

  1. Ultrafast runs GPT-5.6 Sol at 14x standard speed, per OpenAI.
  2. Peak output: up to 750 tokens per second.
  3. OpenAI says you no longer need a smaller model to get real-time speed.

What happened

OpenAI rolled out a new mode called Ultrafast for GPT-5.6 Sol, which the company describes as its latest and most powerful model. The pitch is simple: same brain, much less waiting.

OpenAI says Ultrafast works at 14x the speed of standard processing. The headline throughput is up to 750 output tokens per second. Tokens are the small chunks of text a language model produces when it answers you.

The company's own framing is the interesting part. In its blog post, OpenAI said that until now, real-time speed usually meant picking a smaller or more specialized model. Ultrafast, it argues, points in another direction: more useful work per second.

OpenAI also suggests the mode can be deployed across corporate workflows, with incident response and customer service and support called out as prominent examples.

Competitors are in the same lane. Anthropic's Claude already has a fast mode, though per TechCrunch's reporting it does not reach the speed OpenAI is claiming here.

Note the sourcing: the 14x and 750 figures come from OpenAI itself. There is no independent benchmark on the table yet, so treat them as the vendor's number on the vendor's slide.

Why it matters

Speed has been the quiet tax on frontier models. The biggest ones are the most capable, and also the ones you stare at while a cursor blinks. The usual workaround was to downgrade to a smaller model and accept a dumber answer.

OpenAI is claiming you can skip that trade. If it holds up, the same flagship model that does the hard reasoning can also keep pace in situations where a human is waiting live, like a support chat or an outage in the middle of the night.

That is the real shift. Latency decides where a model can be used at all. A brilliant answer that arrives after the customer has left is just an expensive diary entry.

Real-time speed no longer has to mean a smaller model, OpenAI says.

There is also a competitive signal. Fast modes are becoming a standard menu item, and the fight is now about who can be quick without getting sloppy. Whether Ultrafast preserves quality at full tilt is the question to watch, since OpenAI's pitch is exactly that it does.

Meanwhile, the human reaction time that everyone politely ignored is now officially the slowest component in most workflows. Congratulations on your promotion to bottleneck.

What to do with it

  • Check your latency pain.List the workflows where waiting on a model actually hurts, such as live support or incident triage, before assuming you need more speed.
  • Test, don't trust.OpenAI's 14x is a vendor claim, so run your own prompts and time them against standard mode.
  • Compare with the rival.Claude has a fast mode too, so benchmark both on your real tasks instead of the marketing.
  • Watch the quality.Speed that degrades answers is a bug, so spot-check outputs on anything customer-facing.

Ciao, sapiens. Blink slowly. The model already finished your sentence.