How do I reduce costs when using AI?
At a small business meeting up in Moreton Bay last Thursday, someone asked the best question I’ve heard about AI all year.
He runs a consulting business, where the work product is documents — a lot of them, for clients. He’d been using Cowork, thought the productivity was outstanding, and then the bill arrived.
“How do you manage costs in there?”
He put it to the panel, and got a different angle from each of them. Between them they covered a fair bit — and the rest is documented, it costs nothing, and most of it lives on a settings page that hardly anyone has opened.
Six things that work, roughly in the order I’d do them.
First: why the bill is bigger than the subscription
Two things are going on here, and the tips make more sense once you know about both.
Your subscription is not your bill. The paid Claude plans are subscriptions — Pro is US$17 a month on annual billing or US$20 monthly, and Max starts at US$100 — and that’s the number you think you’ve committed to. But when you reach your plan’s limit you’re offered usage credits so you can keep working; those are billed at standard API rates, charged separately, and they turn up on your bill as additional charges. Auto-reload then tops your balance up whenever it falls below a threshold you’ve set. So you have a productive day, pass the limit without noticing because nothing stops you, and auto-reload quietly buys more. The daily redemption limit is US$2,000, which gives you some idea of how far this can run unwatched.
A long session costs more than it looks like it should. Every message you send carries the whole conversation with it, so the history is re-sent and re-counted on every single turn; the cost doesn’t climb steadily, it curves upward. Twenty turns adding a thousand tokens each isn’t twenty thousand tokens of input — it’s nearer two hundred thousand, because turn twenty is carrying turns one to nineteen along with it.
None of this is hidden, it’s just off to one side while you’re busy working.
Tip 1: Turn the meter on before you change anything
You can’t manage what you’ve never looked at, and hardly anyone has looked.
In Claude it’s all under Settings → Usage: a dashboard showing real-time consumption and your month-to-date credit costs, a monthly spending cap, and the auto-reload switch. Open it before you change a single habit, so you know what normal looks like for you.
If you use OpenClaw, /status will show you the session’s context use and cost as you go. In Microsoft Copilot — one of the panellists works with it and described this — an administrator can set spending limits at organisation and individual user level, with alerts at whatever threshold they choose.
Different vendors, but the same shape of answer: every one of these tools has a meter and a cap, the defaults are generous, and the controls stay switched off until somebody turns them on.
That matters more once you have a team, because four people each working away in their own account means nobody is looking at the total until someone queries the card.
Tip 2: Make auto-reload a decision, not a default
Here’s how I’ve got mine set up — four tools, four different choices, and the same logic each time: the more the work touches someone else’s information, the tighter the control.
Claude Cowork, Max plan, auto-reload off. It tells me when I’ve exceeded my limit and I decide there and then how much to put on. Cowork is where client work happens, so I’d rather be interrupted than surprised.
OpenClaw, auto-reload on at US$250, which takes me a good while to get through. This is my own build and experimentation work, the stakes are lower, and I don’t want to be stopped mid-job.
DeepSeek, US$200 loaded, and I still haven’t had to reload it.
ChatGPT, the $20 plan, mostly for producing images.
An opinion you’re welcome to disagree with: leaving auto-reload on for the tool that touches your clients’ work is a mistake, not because the money is large, but because it removes the one moment where you’d otherwise stop and ask whether the last four hours were worth what they cost. That pause is doing most of the work.
Tip 3: Ask for the content, not the finished thing
This is the change that has saved me the most, so it’s worth showing properly. Say I need a slide deck.
One of the panellists gave a version of this: get it to build one slide first, check the layout and the format, and only then ask for the rest, because every revision rebuilds all of them. That’s sound advice and worth doing.
I go a step further and don’t ask for slides at all until the content is settled.
Ask for the deck’s content as plain text, in the chat — the argument, the structure, what belongs on each slide — then read it properly and argue with it. This is the cheap part, and it’s the part that actually matters.
Once the content is right, one carefully written prompt to turn it into slide content; a good prompt gets me about 98% of the way there on a first draft.
Copy that into my own template, on my own machine — my fonts, my colours, my layout, all set up once, years ago.
The model never sees a backdrop; it doesn’t choose a font, pick a colour or lay anything out.
The reason this saves so much is that output is priced at exactly five times input on every current Claude model. Ask for the deck again and you pay to write every slide over, and then both versions sit in the conversation history being re-sent on every turn after that. Change your mind about one heading and you’ve paid for the whole deck twice.
If your business produces documents for clients, this is where the money goes — not on the thinking, but on the re-formatting. Do the decorating once, yourself, in a file you keep and reuse.
Tip 4: Work in short, finished sessions
Providers cache repeated context, which softens that curve in dollars; on Claude a cache hit costs a tenth of the normal input rate. But caching has a short memory, measured in minutes to about an hour, so it rewards you for finishing a job in one go and quietly stops helping when you leave a big session open and drift back to it across the afternoon.
So work one task at a time and finish it, then start a new chat for the next thing rather than running one endless thread carrying three unrelated jobs.
Anthropic’s own advice maps onto the same mechanism: keep project instructions short, since they’re re-sent every turn; turn off tools you aren’t using, like web search; drop the effort level for routine work; turn off extended thinking when the task doesn’t need it; and put reference material in a project rather than pasting it into the chat.
Tip 5: Match the model to the job
The spread is wider than most people realise. Claude Haiku 4.5 is US$1 per million tokens in and US$5 out, while Opus 5 is US$5 and US$25, so the same question put to a different model can cost five times as much.
Most admin work — sorting, summarising, extracting, reformatting — doesn’t need the most capable model available, so save that for the jobs where being wrong is expensive and let the cheaper one do the sorting.
It’s worth asking for shorter answers too. Output is the expensive direction, so “give me this in five bullet points” is a genuine saving rather than just a style preference.
Tip 6: Know where the cheap models don’t belong
People ask about DeepSeek, so here’s my rule, which is a good deal shorter than the explanation: I use DeepSeek more than Claude for APIs, but not when it comes to my clients’ information.
The appeal is obvious from the rate card. DeepSeek’s Flash model runs US$0.30 per million tokens in and US$1.20 out at peak, and half that off-peak; peak is a narrow window, so most of the time you’re paying the lower rate. For bulk work where the task is simple and the output gets checked, that gap adds up to real money.
The caution is real too, and worth stating accurately rather than dramatically. In February 2025 the Secretary of Home Affairs issued PSPF Direction 001-2025, requiring Australian Government entities to prevent the use of DeepSeek products and remove existing instances from government systems. That direction binds government entities; it is not a general ban on private businesses using it — although if you hold Commonwealth sensitive information under contract, check what your deed obliges you to do.
What it should prompt is the question worth asking of any tool before you send it something that isn’t yours: what are the terms of the plan I’m on, what’s retained, and is any of it used for training?
Cheap is a perfectly good reason to use something for the work that doesn’t matter much, and a poor reason to use it for the work that does.
If your AI spend has started to move and nobody can quite say why, that’s usually a sign the tools arrived before the decisions did. An AI Readiness Day is a day in your business working out which handful of jobs are actually worth automating, what they should cost to run, and where the guardrails go — before the spend starts rather than after. $3,500 + GST, with a credit towards a larger project if you commission one within 30 days.
Sources, all checked 14 September 2026: Anthropic Help Centre (usage credits, and usage and length limits); Anthropic pricing; OpenClaw documentation; DeepSeek API pricing; PSPF Direction 001-2025, Department of Home Affairs; Microsoft Learn (usage-based billing for Copilot Credits). Prices here are US dollars and move often — check the current rate card before budgeting off them.
