Spending Fewer Tokens

Talking to an AI costs money in proportion to how much talking there is. Nearly all of that cost is invisible — it is not what you type, it is the conversation being re-sent, over and over, behind you. This page is the short list of habits that cut the bill, ordered by how much they actually save.

Why a conversation costs what it does

Every time the assistant calls a CAD operation, that is not the end of the exchange — the result comes back and the whole conversation is sent to the model again so it can decide what to do next. A request that takes twelve operations is therefore not one exchange but about a dozen, and each one carries everything that came before it.

The effect is lopsided in a way that surprises people. Measured on this product, across a window of 25 conversations: 4.16 million tokens sent to the model, against 58 thousand written back. That is roughly 72 to 1. What you are paying for is almost entirely context being re-sent, not answers being written.

Which gives the rule everything below follows from: fewer rounds, shorter history, and — best of all — no model at all.

1. Record a macro and reuse it

This is the biggest single saving available to you, and it is not close.

A macro run calls no AI model whatsoever. Every step of a run goes straight to your CAD system through the same direct dispatch a single operation uses; there is no model reachable from inside a run. So a job you talk through today at full price can be run tomorrow, and every day after that, for nothing.

The pattern to adopt: the first time you do a repeatable job, do it with the recorder running. Talk it through as normal — that part costs what it costs. Stop the recording, accept the parameters you are offered, save it. From then on the job is a button. See Macros.

It also scales the right way. Running the macro over ninety parts costs exactly what running it over one part costs: nothing. Doing the same ninety parts by conversation costs ninety conversations, each one longer than the last.

2. Use a variation template for a family of sizes

Same reasoning. Generating eight named instances from one master is one operation, not eight conversations — the values are written, read back and verified by the CAD system directly. The only model time is the one request that starts it.

If you find yourself saying “now make it 120 wide… now 140… now 160…”, stop and capture a template instead. See Variations.

3. Pick the Target before you start

If more than one CAD system is connected and your request does not say which you mean, CoPilot will not guess — it stops and asks. That is the right behaviour, but answering the question is a whole extra round with the full context attached.

Set the Target box on the chat card to the session you want and the question never comes up. See The Desktop Agent.

4. Ask for the whole job in one message

Drip-feeding costs you. Compare:

Create a 120 x 80 x 15 mm plate

…then, after it finishes, “add four Ø8 holes”, then “now fillet the corners”, then “now save it”. Four separate turns, each re-sending everything before it. Against:

Create a 120 x 80 x 15 mm plate, add four Ø8 through holes 15 mm in from each corner, fillet the vertical corner edges at R3, and save it as C:\parts\plate.sldprt

One turn. The same work, and the assistant can also issue independent operations together rather than one per round.

This pulls against the advice in Chat Basics to work one feature at a time on complicated models — and that advice wins when the geometry is hard, because a wrong twenty-feature run costs far more than a few extra turns. The rule of thumb: batch the routine, step through the difficult.

5. Start a new conversation for a new job

A conversation carries its whole history into every round. Once you have finished with the bracket and moved on to the housing, the bracket is still being re-sent — paid for, and adding nothing.

Clear the conversation (/clear in the chat, or start a new one) whenever the subject genuinely changes. There is also /compact, which shortens a long conversation in place while keeping what matters.

6. Reach for a skill when the task varies, a macro when it repeats

A skill is a standing instruction. Writing “all drawings are A3, third angle, one decimal place, centre marks on every hole” once, as a skill, is cheaper than typing it into every drawing request — and far cheaper than a round of correction when you forget.

The two work together: use a skill for the rules that vary in application, a macro for the steps that do not vary at all.

7. Ask read-only questions cheaply

Questions about the model are usually one or two operations, so they are inexpensive — but they are not free, and they are the easiest thing to ask carelessly. Two small habits:

Ask for several facts at once. “Give me the volume, the bounding box and the number of faces” is one round; three separate questions are three.
Say which measurement you want. A precise question is answered with a precise tool call; a vague one may be answered with several exploratory ones.

8. Know where the meter is, and whose it is

Two different bills are possible, and it is worth knowing which one you are on.

Your organization's plan. Usage counts against the allowance on your seat.
Your own AI key (BYOK). Add a key on the organization portal under My AI Key and your chats run on your own provider account, billed by that provider directly. On a trial this is not optional — a trial user must supply a key before chat runs at all.

Either way, the organization portal's licence page shows the last 30 days of usage, split into what ran on the plan's allowance and what ran on your own key. In the chat itself, /usage reports this month’s tokens and cost, and /context shows how full the current conversation’s context window is — the chat also carries a gauge for it: green under 70 %, amber to 90 %, red above. See Organization Portal.

Different models cost very different amounts per token for the same job, and the AI drop-down above the input box lets you switch between your own API key and a coding agent without leaving the chat. It is worth trying a smaller model on routine work and keeping the larger one for the modelling that actually needs it — but verify the results either way, because a cheaper model is a cheaper model. See The Desktop Agent and Verifying Results.

The short version

Habit	Why it saves
Record a macro for anything you will do twice	A run calls no model at all — the biggest saving by a wide margin
Use a variation template for a family of sizes	N instances is one operation, not N conversations
Pick the Target before you start	Removes a whole round spent establishing which CAD you meant
Ask for the whole routine job in one message	Each extra turn re-sends the entire conversation
Start a new conversation for a new subject	Stops paying to re-send work you have finished with
Write house rules as a skill	Said once rather than in every request — and no correction round
Group read-only questions	Three facts in one round instead of three rounds