OpenAI Ultrafast: 8x Speed at 6x the Price
OpenAI Ultrafast promises up to 8x faster Codex output for about 6x the API price. Here is the arithmetic on when paying for latency pays back.
Speed now has a price list, and the numbers are close to a wash. OpenAI Ultrafast, a premium tier announced at DevDay on 29 September 2026, generates up to eight times faster in Codex and up to six times faster in the API, for about six times the standard API price. It runs on Pro 500 and Enterprise plans for GPT-6 Astra, with a GPT-6.1 Sol version described as coming. When the price multiplier roughly matches the speed multiplier, the question stops being whether it is fast and becomes whether latency is the thing you are actually short of.
What OpenAI published about Ultrafast
Per BGR's DevDay roundup, Ultrafast runs up to eight times faster in Codex and up to six times faster in the API, at six times the standard API price, currently on Pro 500 and Enterprise for GPT-6 Astra.
One caution on throughput figures. digit.in reports up to 300 tokens per second in Codex, while BGR puts the forthcoming Sol version at up to 350 tokens per second. Those are not the same claim, and OpenAI has not published a consolidated table, so treat any specific tokens-per-second number as unconfirmed until it appears in the official pricing docs.
When 6x the price for 6x the speed is worth it
Paying a linear premium for linear speed only makes sense when something other than tokens is the expensive part. Three cases where it genuinely does:
A human is waiting. If an engineer sits idle while a coding agent grinds, their hour costs far more than the token delta, and an eight times faster Codex turn changes how the day feels.
The work is interactive. Anything where the user abandons the task at ten seconds has a revenue cliff that token pricing does not capture.
A run has a wall-clock deadline. A nightly job that must finish before the business opens is a scheduling problem, and speed is the only lever.
And the cases where it is wasted money: batch work nobody is watching, background summarisation, anything queued. There, six times the price buys you a result that sits in a queue slightly sooner.
The cheaper lever most teams have not pulled
The same event shipped GPT-6.1 Sol at roughly a fifth of Astra's token cost while claiming close to Astra-level performance on agentic coding and computer use. For a lot of workloads that is the better trade: a model that costs 20 percent as much beats a tier that costs 600 percent as much, unless latency specifically is your bottleneck. Our write-up of the GPT-6.1 Sol launch has the detail.
Before buying speed, it is worth confirming where your time actually goes. We worked through the general version of this trade in whether a faster AI coding model is worth the price, and the cost side compounds in ways people underestimate, as in why long context costs more.
How to decide in an afternoon
Latency is the one dimension of model performance that people buy on feel rather than measurement, which is exactly why a six times premium is easy to waste. Fifteen minutes of timing data settles it.
Measure the wall-clock time of your ten most common agent runs. If none exceed the patience of whoever waits on them, stop here.
For the ones that do, multiply the minutes saved by the loaded hourly cost of the person waiting, then compare against the six times token premium on that specific workload.
Check plan eligibility before budgeting. This is a Pro 500 and Enterprise feature today, not an account-wide setting.
Pricing tiers move fast enough that comparison is a standing task rather than a one-off, which is why we keep a method for comparing AI API pricing across providers and track releases through our approach to keeping up with AI news.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
