Start with the right equation
AI cost per task is people time plus attempts multiplied by cost per attempt. Most cost-cutting work only attacks cost per attempt, by switching to a cheaper model. That can raise the bill if the cheaper model needs more retries, because every retry costs tokens and staff time. Work through the levers below in order, and measure the result as cost per accepted task, not price per token.
1. Measure cost per accepted task
Count, for each workflow, the model spend plus the staff time spent fixing and re-asking, divided by the outputs that passed your acceptance criteria. Without this number every other lever is a guess, and a cheaper model can look like a win while the true cost rises. Start with your top three workflows by spend.
2. Cut retries before you touch models
If a task takes four or five attempts, the cheapest saving is the attempts you remove. Retries come from inconsistent prompts, missing context and no shared definition of “done”. Capture how your best people get a task accepted first time, then reuse it for everyone. That is the idea behind NeoAI's distilled recipes. You can start by hand: write down the steps, context and checks for your five most repeated tasks.
3. Match each step to the cheapest model that passes
Use a small model for routine steps such as parsing, extraction and drafting, and keep the frontier model for steps that need judgment. The prize is large: on list prices read on 25 Oct 2026, a mid-priced model costs about 20x a small model in the same family Public list prices. See the dated price table. Decide per step, not per application, and test each choice against your acceptance criteria, not an average benchmark.
4. Trim context
Input tokens are billed on every call, and agents resend their whole history on each step. Send only the sections a step needs, summarise tool outputs before they enter the running history, and cap loops. In our simulated replay, shortening what the frontier model sees was a main reason token use fell 65–72% Simulated. That is a simulation, not a promised result.
5. Cache repeated prefixes
System prompts, tool schemas and shared instructions repeat on every call, and providers bill cache reads at a fraction of base input. On Anthropic's list, Claude Haiku 5.5 cache reads are $0.01 per million tokens against $0.10 base, and Claude Opus 5.5 is $0.20 against $4.00 Public list prices. Keep stable content at the start of the prompt and variable content at the end so the prefix can be reused.
6. Batch work that can wait
Anthropic's Batch API gives a 50 percent discount on both input and output tokens for asynchronous work Public list prices. Overnight classification, back-fills and evaluation runs rarely need a live answer. Move them to batch before optimising anything else on those workflows.
7. Gate quality, and escalate instead of failing
Every saving above is only a saving if quality holds. Define acceptance criteria per task, run a fixed evaluation set before and after each change, and add a confidence or quality check on cheaper steps. When a cheap model is unsure, move that step up a tier automatically rather than returning an error or a poor answer. This is how NeoAI's smart delegation works.
Common mistakes
- Switching everything to the cheapest model. Retries and rework can cost more than the tokens saved.
- Judging by an average benchmark. Your tasks and your acceptance criteria are what count.
- Measuring spend, not outcomes. Falling spend with falling acceptance is not a saving.
- Ignoring staff time. Hours spent fixing outputs are part of the cost.
- Optimising once. Models and prices change, so re-tune on a schedule.
Where a tool helps
You can run levers 1, 4, 5 and 6 yourself. Levers 2, 3 and 7 are where it is hardest to do by hand, because they need recipes learned from your accepted work and a per-step decision with automatic escalation. If you spend $20k a month or more on LLMs, a free audit sizes what those levers are worth on your data. Pricing is gainshare: no measured savings, no fee.