OpenAI Ultrafast: Faster Output, Not Faster Everything

When we ask an AI coding tool for help, waiting for the answer can interrupt the work. OpenAI calls Ultrafast its premium speed tier, with claims aimed at how quickly output is generated in Codex and the API. That makes the promise attractive—and makes the distinction between faster output and a faster finished job worth keeping.

If we are ready to continue but still watching an answer appear, faster output could shorten that pause. If the harder part is checking the answer or making it useful, quicker generation alone may leave much of the job untouched.

Before judging the value of a premium tier, we need to separate two observations: when the output arrives and when we have something we can use. What, exactly, do OpenAI’s announced numbers measure?

OpenAI says Ultrafast offers up to 8x faster token generation in Codex, with 300 tokens per second stated alongside that claim. For the API, OpenAI separately says up to 6x. The 300-token figure belongs to Codex; the sentence does not assign it to the API.

“Up to” is part of both claims, not a promise of the same increase on every task. The post gives no benchmark conditions or shared comparison baseline, so these figures cannot support a calculated baseline throughput or a prediction of our total task time.

The complete sentence keeps the important boundaries together: the premium tier, the “up to” qualifications, token generation and the two different settings. Reading it intact avoids turning a qualified output-speed claim into a promise about the whole job.

The captured post records that announcement. It is evidence of OpenAI’s statement, not an independent test of Ultrafast.

Our premium speed tier, Ultrafast offers up to 8x faster token generation (300 tokens per second) in Codex and up to 6x in the API.

OpenAI · https://x.com/OpenAI/status/2104993966043320759

The available video evidence is limited to one observed frame of OpenAI’s attached comparison. The full video and audio were not downloaded or transcribed, so that frame cannot support a complete account of the demonstration or a validated comparison of timings.

The useful evidence here is narrower: we can preserve the exact claim and identify who made it. We cannot turn the illustration into measured results or derive another speed ratio from it.

We ask for help, the AI produces output, and then we check whether we can use the result. Tokens are pieces of that generated output—not finished tasks and not necessarily whole words. Token-generation speed describes the production step.

That distinction does not tell us how Ultrafast works internally, how good its answers are or how long checking will take. It simply explains why an increase in output speed does not establish an equivalent reduction in the time needed to finish the job.

We would first check current official information for access and cost. This post gives no price, eligibility, rollout scope or activation instructions, so it cannot tell us whether we can use Ultrafast or what we would pay.

If we confirm access, we could compare the same task with and without the tier, recording output arrival separately from completion of a checked, usable result. Keeping the task and checking criteria comparable would make the observations more useful, without turning them into a controlled benchmark.

This is a proposed test, not one we have run. Its purpose is to find out whether the waiting Ultrafast targets is the waiting that matters in our work.

Ultrafast’s appeal is specific: potentially less waiting while output is generated. Whether that becomes a worthwhile benefit depends on access, cost and how much that pause holds up our task—none of which this announcement settles.

The prompt below uses OpenAI’s exact passage to separate the claims from the unproven payoff and propose a comparison we could make. It asks for a decision note, not invented benchmark results.

Faster output is the promise. Usable work is the test.

Altior · editorial interpretation