Why it matters
Model choice is often the largest lever on AI unit cost. Teams tend to default to the most capable model for every task, including routine work a smaller model handles well.How nOps finds it
nOps reads the last 30 days of AI billing for each owner and model: a person or workload on AWS Bedrock and Claude Enterprise, an API key on the Anthropic API, and a project on the OpenAI API. On Bedrock and Claude Enterprise the only move suggested is Claude Opus to Claude Sonnet. On the Anthropic and OpenAI APIs the target is exactly one tier down in the same family, never an older generation. The target must be a model your organization spent at least $100 on in the window, the owner must have spent at least $25 on the current model, and the target must have a price for every token type the owner used (input, output, cache reads and writes). The saving is each token type’s cost times the difference between the two models’ rates, using the rates on your own bills, and must be at least $25 a month. When a newer model in the same tier also saves money, nOps recommends that version upgrade instead.Before you start
The ability to change the model the workload, API key or user calls, and a way to check output quality on the cheaper model.How to implement
- Find the owner and current model named in the recommendation.
- Run a representative sample of real prompts against the suggested model and compare the answers, tool calls and latency with today’s model.
- If quality is acceptable, change the model ID in the application configuration, the gateway or the user’s tool settings. Keep the current model as a fallback for the tasks that need it.
- For Claude Enterprise and Claude Code users, ask the person to make the cheaper model their default and switch up only for hard tasks.
- Watch spend and quality for one to two weeks after the change.
Risks and considerations
- A lower tier can give weaker answers, follow complex instructions less reliably, or handle tools and long context less well. Test before moving production traffic.
- The estimate assumes all of this owner’s work on the model moves. Moving only part of it saves proportionally less.
- To roll back, switch the model ID back.
FAQs
Why does nOps only suggest models we already use?
Why does nOps only suggest models we already use?
nOps prices the move at the rates on your own bills, so the target must be a model your organization already runs with at least $100 of spend in the window.
Is this a guaranteed saving?
Is this a guaranteed saving?
No. The figure assumes the same work runs on the cheaper model. The quality impact is something you need to check.