Five Ways to Cut LLM Costs Without Changing a Line of Code
Batch pricing, prompt caching, model selection, idle seats and unit cost tracking. Five levers every FinOps team should check first.
When AI spend climbs, the first instinct is often to ask engineering to rewrite things. That is slow, it competes with the product roadmap, and it rarely survives the next sprint. In practice, a large share of LLM cost optimization comes from decisions about pricing models, model selection, and license hygiene. Most of them need a configuration change or a procurement decision, not a refactor.
Here are five levers every FinOps practitioner should check first.
1. Use batch pricing for work that is not time-sensitive
Major model providers, including OpenAI and Anthropic, offer batch APIs that process requests asynchronously at roughly half the standard token price. Nightly summarization, document classification, evaluation runs, and data enrichment rarely need a response in seconds.
What to look for: high-volume workloads with steady, predictable traffic and no user waiting on the other end. These are your best batch candidates.
2. Turn on prompt caching for repeated context
Many applications send the same system prompt, tool definitions, or reference documents with every request. Prompt caching lets the provider bill those repeated input tokens at a reduced rate. For chat assistants, agents, and RAG pipelines with long fixed instructions, the effect on unit cost can be meaningful.
What to look for: a high ratio of input to output tokens on a workload, combined with long, stable prompts.
3. Match the model to the task
Not every request needs a frontier model. Classification, extraction, routing, and short reformatting tasks often run well on a smaller, cheaper model. The price gap between model tiers can be large, so moving a high-volume, low-complexity task down one tier can change the unit economics of a feature.
What to look for: simple tasks running on premium models. Always review the quality impact of each swap with the owning team before you act.
4. Reclaim idle seats
Coding assistants and AI chat tools are often bought in bulk during a rollout. Months later, a meaningful number of those seats show little or no activity. This is the AI version of the idle instance problem, and it is usually the fastest saving to capture.
What to look for: licensed users with no usage in the current period. Billed seats that are not assigned to anyone at all.
5. Watch cost per request, not just total spend
Total spend can rise for good reasons, such as adoption. Cost per request rising is a different signal. It often points to longer contexts, more retries, or a quiet switch to a more expensive model. Tracking unit cost by team and model turns a vague concern into a specific conversation.
What to look for: a team or agent whose cost per request jumps while request volume stays flat.
Turning levers into a monthly habit
The hard part is not knowing these levers exist. It is finding where each one applies, quantifying it, and routing it to the right owner every month.
KeepAir does that work automatically. It analyzes usage across 15+ AI vendors and produces ranked savings suggestions, each with an estimated monthly dollar figure: caching opportunities, model swaps with the quality impact shown, batch pricing candidates, idle seats, and cost trends. Every suggestion is grouped by scope, from the whole organization down to a single team, user, or agent, so it reaches the person who can act on it.
Setup takes one API key per vendor and no code changes. KeepAir is free to use.
Find your first savings this week.
Also published on LinkedIn. Follow KeepAir on LinkedIn