THE VERDICT
● Standard for almost everything. Batch work, background jobs, anything asynchronous.
● Fast mode only where a person is waiting, or where latency compounds across many steps.
● The test: can you name what the saved time is worth? If not, you are paying double for nothing.
Where Fast mode earns it
| Scenario | Worth it? |
|---|
| A user is watching a spinner | Yes. Perceived speed is product quality |
| Agent loop with many sequential steps | Often. Latency compounds across dependent steps |
| Interactive coding session | Maybe. Depends whether you context-switch while waiting |
| Overnight batch job | No. Finishing at 3am instead of 5am is worth nothing |
| Bulk classification or extraction | No. Use a cheaper model, not a faster one |
THE COMPOUNDING CASE IS THE REAL ONE
A single call finishing two seconds sooner is invisible. An agent making forty sequential calls finishing eighty seconds sooner is a different experience.
That is the same argument NVIDIA made for splitting decode onto dedicated silicon — agent loops are serial, and latency per step is what people actually feel.
The cheaper thing to try first
Astra now lets you change reasoning effort without invalidating the prompt cache. Running lower effort on simple steps and raising it only where needed will usually cut wall-clock time more than doubling the rate would, and it costs less rather than more.
Try that before paying for Fast mode. If it is still too slow afterwards, the case for Fast is genuine.
FAQ
How much faster is Fast mode?
Twice the standard speed at twice the standard price, per OpenAI. Per-million pricing was not published at launch, so verify the base rate before calculating anything.
Does Fast mode change output quality?
It is presented as a speed option rather than a different model. Test on your own work rather than assuming equivalence.
What should I try instead?
Varying reasoning effort per step, which is now possible without losing the prompt cache. Cheaper and often more effective than paying for speed.