How to reduce your AI API costs

How you are actually billed

Providers charge per token, priced separately for input and output, at rates that vary by model. Output tokens usually cost several times input tokens. That single fact explains most of what follows.

We are not printing a price table. Provider prices change, and a stale number here would quietly make your planning wrong. Get current figures from your provider's own pricing page, then record them per profile and model in SideNote Pro so the local estimates mean something.

The levers, roughly in order of impact

1. Use a cheaper model for the majority of your work

Price differences between model tiers are large, and most everyday requests are fluency and comprehension tasks that do not need the top tier. If you are running everything through your most capable model, this one change usually dominates every other saving combined. See choosing the right AI model.

2. Do not pay for reasoning you do not need

Reasoning effort generates tokens before your answer starts, and those are billed. On a task with no hidden structure that is pure cost. Keep everyday profiles on Provider default and raise effort only on the deep profile. See reasoning effort.

3. Control conversation context

Every turn re-sends accumulated context, so a long thread charges you for the same history again and again. Starting a new thread for a new topic is the highest-value habit in this whole list, and it costs one keystroke.

4. Ask for shorter answers

Output tokens are the expensive ones. *In five bullets* costs a fraction of an unbounded essay, and is usually more useful. Unbounded prompts produce unbounded answers.

5. Be deliberate with attachments

A large document is a large input on every subsequent turn of that thread, not just the first. Ask what you need, then start a fresh thread rather than carrying the document through an unrelated conversation.

6. Regenerate less, edit more

Regenerating repeats the full input cost for another sample. When the problem is the question, editing and resending is both cheaper and more likely to work than a third roll of the dice.

Compaction is not free

When a thread exceeds its context budget, SideNote Pro asks your configured model to summarize the older messages. That is a real provider request with real tokens, and it happens on the profile you have selected, at whatever reasoning effort it is set to.

Compaction is a good deal on balance, since it stops you re-sending an entire long thread every turn. But frequent compaction is a signal that threads are running too long. Splitting them is cheaper than summarising them repeatedly.

Local models change the arithmetic

A local model through Ollama or LM Studio costs nothing per token. For high-volume repetitive work such as rewriting, summarizing and translating, moving that to a local profile can remove a large share of your bill outright.

The cost moves rather than vanishing: it becomes your hardware and your time. But for the fluency tasks that dominate most people's volume, a small local model is often entirely adequate.

Guard rails worth setting today

1. Set a hard spending cap in your provider dashboard. This is the only authoritative limit. Do it before anything else.
2. Set the optional daily estimate limit in SideNote Pro as a local seatbelt. It works from local estimates and cannot see other applications using the same key.
3. Record your model prices per profile in USD per million tokens, so local estimates are meaningful.
4. Look at the ledger weekly for the first month. Which provider and model combinations dominate, and were they doing work that needed them?

Keep it in perspective

The point of watching cost is not to minimise it. An answer that saves you an hour is worth more than the cents it cost, and optimising a five-dollar month down to four is not where your attention belongs.

The point is that spending should correspond to value. Paying top-tier reasoning rates to rewrite a sentence is not thrift versus quality, it is just waste. Fix that, and stop worrying about the rest.