How deadlines become discounts: SLA tiers explained
The deadline you set is the main lever you have on what a batch job costs. The wider the window, the cheaper the work.
A deadline is an option to wait
When you submit a job with a deadline, you are handing the scheduler an option: the right to wait for a better price up to that point in time. Compute prices move — across regions, across providers, across the hours of the day. Every hour of slack you grant is another chance to catch a dip.
This is why the same job can be priced very differently depending only on when you need it. The work is identical; what changes is how much room the scheduler has to find a cheap moment to do it.
How tiers turn patience into price
SLA tiers package this into a small set of named completion windows, longest (cheapest) to shortest (priciest):
| Tier | Window | Behavior | Good for |
|---|---|---|---|
| Remnant | < 7 days | Maximum patience; waits for deep troughs and idle gluts. | Backlogs, reprocessing, anything not time-bound. |
| Standard | < 24 hours | Aggressive price-waiting, cross-region spot, heavy batching. | Nightly enrichment — the everyday default. |
| Priority | < 6 hours | Waits for moderate dips, prefers warm workers. | Same-business-day batches. |
| Rush | < 1 hour | Takes near-current capacity, minimal waiting. | "I need it this hour." |
Choosing the right window
Pick the widest deadline the work can genuinely tolerate, not the tightest you can imagine wanting. Teams routinely over-specify urgency out of habit and pay for it. Ask: when does this output actually need to exist? If a catalog refresh feeds a weekly report, a 7-day window costs far less than a 1-hour one and changes nothing about the outcome.
You can also mix tiers across a pipeline: run the bulk of a backlog on the cheapest tier, and reserve a faster tier only for the slice that is genuinely time-sensitive.
Predictable price, hidden volatility
A good platform shields you from the underlying volatility. You see a fixed, predictable rate per tier — a published rate card you can plan against — while the platform absorbs the moment-to-moment swings of the spot market. You get a stable price and the discount; the platform carries the risk of finding capacity below it.
How this compares, and where it comes from
The idea that patience is worth money is not novel — the interactive vendors sell a version of it. The notable part is how coarse that version is. OpenAI's Batch API prices a flat 50% discount, and its completion window can only be set to 24 hours: a single point on the curve, take it or leave it.
Tiers are what that looks like when the window becomes a dial rather than a constant. The reason a wider dial reaches a lower floor is the subject of a specific piece of systems research: with a deadline, a scheduler can run on interruptible capacity and hold on-demand in reserve, switching only when the job falls behind the pace its deadline demands. That policy is the contribution of Can't Be Late (NSDI 2024), and it is what a tier is quietly buying you.
Frequently asked questions
Why is a longer deadline cheaper?
A longer deadline gives the scheduler more time to wait for cheap, interruptible capacity and to shop across regions and providers for a price dip. More waiting room means a lower achievable price.
Which SLA tier should I choose?
The widest window the work can truly tolerate. If no one is blocked on the result, prefer the longest tier; reserve faster tiers only for the genuinely time-sensitive slice of a pipeline.
Related guides
Spot and remnant GPUs are the same hardware sold cheap, with the catch that it can be reclaimed. How it works, and why batch can use it when interactive cannot.
Batch AI inference runs large volumes of requests against a deadline instead of in real time. How it differs from interactive inference, and why it costs less.
Prompt caching reuses the computed state of a repeated prefix so you stop paying full price for it. How it works, why batch benefits most, and how it is billed.
If your AI work is a pipeline, submitting one request at a time discards the structure that would let it be scheduled well. What the research says that costs.