Why it matters
Fast options are often turned on once for convenience and never revisited, and the premium applies to every request, including background and batch work where nobody is waiting.How nOps finds it
For the OpenAI API, nOps reads the last 30 days of spend per project and model on each premium tier. It measures how much more the tier costs than standard from your organization’s bills, or uses a default (priority 2 times, ultrafast 6 times) when there is not enough data. A tier must cost at least 1.2 times standard. The saving is what that spend costs over the standard price, assuming 50% of the traffic moves, and must be at least $25 a month. For Cursor, nOps reads the last 30 days of each user's fast model requests. A user needs at least 30 fast requests, and the fast variant must cost at least 1.2 times the standard version per token at your organization's own rates. The saving is the whole difference for that user's fast usage, at least $25 a month.Before you start
The ability to set the processing tier per request or per workload, or to change the model selection in Cursor.How to implement
- Decide which calls really need the fastest response, such as interactive features where a user waits.
- OpenAI API: send everything else with the standard service tier (set
service_tiertodefault, or leave it unset), or use Flex or the Batch API for work that can wait. - Cursor: ask the user to pick the standard version of the model instead of the fast variant for routine work, and keep fast mode for when speed matters.
- Watch response times on the paths you changed for a week.
Risks and considerations
- Standard processing is slower and can queue at busy times. Keep latency-critical paths on the premium tier.
- For OpenAI the estimate assumes half the premium traffic moves. Your real saving depends on how much can wait.
- To roll back, set the tier or model back.
FAQs
How much traffic does nOps assume moves?
How much traffic does nOps assume moves?
For OpenAI, half of the premium-tier traffic. For Cursor, all of the user’s fast-variant usage, because the user chooses the variant per request.
Related
- Use Bedrock Batch Inference for Latency-Tolerant Workloads
- Move Latency-Tolerant Azure OpenAI Workloads to Batch
- Supported recommendations