SideNote ProHow to reduce your AI API costs
Metered billing rewards attention in a way subscriptions do not. The good news is that most overspending comes from a few habits, and each of them has a direct fix.
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. Most ways to cut cost follow from that single fact.
Provider prices change, so use your provider's current pricing page rather than a static price table. Record those rates 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, and for one dated, worked example of what an everyday model costs per request, the GPT-6 Luna setup.
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. Where your everyday model accepts it, None is what stops paying for it; Provider default only omits the field, so a model that reasons by default carries on doing so. Raise effort on the deep profile instead. 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 outgrows the context window it is being sent on, 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
- Set a hard spending cap in your provider dashboard. This is the only authoritative limit. Do it before anything else.
- 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.
- Record your model prices per profile in USD per million tokens, so local estimates are meaningful.
- Look at the ledger weekly for the first month. Which provider and model combinations dominate, and were they doing work that needed them?
Local cost figures are estimates. SideNote Pro does not query your provider's billing system and cannot reconcile a bill. They are for comparing configurations, not for accounting. See usage and cost tracking.
Settings › About
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.