Why is my AI bill this high and how do I cut it?

Most of an AI bill is the context you send, not the answers you get back — the input side is roughly 90% of the cost. The fix is sending less: a narrower working folder, fresher sessions, and caching the parts you reuse.

Is AI really billed on what I send, not what it writes back?

Both are billed, but they're wildly lopsided. In a typical coding workload the input — everything you feed in as context — is around 90% of the cost, and the output the model writes back is only about 10%. That's why shortening the answer barely helps: you're trimming the small side of the ledger. The money is in what you send.

Won't switching to a cheaper model fix the bill?

It helps, but less than you'd think. In one measured case, the choice of model accounted for only about 30% of total spend — the other 70% was the context being fed in. A cheaper model applied to a bloated context is still paying to process the bloat every request. Fixing what you send often saves more than swapping the model.

Doesn't a good AI coding tool already avoid sending whole files?

The better ones do — which is exactly why the eye-popping "90-something percent saved" figures are worst-case numbers, measured against a naive baseline that dumps entire files every time. If your tool is already selective, your real savings are smaller. The principle still holds: the less irrelevant context rides along on each request, the less you pay.

Can't I just tell the model to \"only use what's relevant\"?

No, and this is the trap. You're billed for the context the moment you send it — before the model has read a word of your instruction. Telling it to ignore the irrelevant 40,000 tokens doesn't refund them; you already paid to transmit them. The saving has to happen before you hit send, by not including them in the first place.

Do I need an expensive laptop to use Claude?

No. Claude runs on Anthropic's servers, not on your machine — the Desktop app is a window onto a model running in the cloud. Any modern laptop that can open a browser and a chat window is enough. The expensive hardware conversation is about running other AI models locally on your own device, which is a niche need most professionals never have.

What is a RAM or memory-price crisis and why does it affect laptop prices?

AI accelerators are built with the same memory and flash-storage chips used in ordinary laptops and desktops. When the largest AI companies buy that supply in bulk, less is left for consumer hardware, and prices rise for everyone downstream. In this cycle a memory kit that cost around $100 a year ago has been selling for five to six times that, and finished laptops are following.

Why is unified memory good for running AI models locally?

On most PCs the graphics chip has its own small pool of memory, separate from system memory — often capped around 32GB — and a model too big for that pool slows down dramatically. Some laptop chips instead share one large memory pool between the processor and graphics, so a 128GB machine can load a very large model entirely into memory and run it quickly. That shared design is why certain laptops became the value pick for local AI.

Should I buy a machine now or wait for prices to fall?

If you only need to use AI, neither — a modest laptop plus a cloud subscription covers you, and you're insulated from hardware prices entirely. If you specifically need to run large models locally (privacy rules, offline work, heavy experimentation), buy the memory you need now rather than betting on a near-term drop; supply pressure is expected to persist, and refurbished high-memory configurations are currently the better value than new.

Which Claude plan do I need?

Start on Pro if you're an individual — it includes the full product, Cowork and Claude Code included, and you'll discover your real usage pattern within a fortnight. Move to Max if you're hitting limits regularly rather than occasionally. Team and Enterprise matter when you need commercial data terms, admin controls, and SSO, not just more capacity.

Why did I hit a limit when I'd barely used Claude Code?

Because the allowance is shared. Claude Code, claude.ai chat, and Cowork all draw on the same bucket — a heavy morning in chat reduces what's left for a coding session that afternoon. Run /usage to see what's actually consuming it.

How do the limits work?

There are two layers. A rolling five-hour window governs short bursts of activity, and a weekly cap governs sustained use, resetting seven days from your first message at 00:00 UTC. Hitting one doesn't necessarily mean hitting the other — a five-hour limit clears in hours, a weekly one doesn't.

What AI costs — Claude for Humans

There are two different bills hiding behind “what does AI cost.” One is the seat someone pays for each month — a purchasing decision made once, then forgotten. The other is usage-based spend, which moves every day depending on how people actually work. It’s the second one that produces the meeting where a number went up and nobody changed anything.

The surprise is always in the same place, and it’s the opposite of where people look. The expensive part isn’t the model thinking, and it isn’t the long answers it writes back. It’s everything you send in — the files, the history, the standing instructions riding along on every single request. Which means the cost question and the context question are the same question wearing different clothes: a tight, deliberate context is both cheaper and better. You rarely have to trade one for the other.

That framing also explains why prompt caching matters more than its name suggests. If you’re going to send the same long instructions or the same big document turn after turn, the expensive work of reading it can happen once and be reused at a fraction of the price. It’s defined below, and it’s most of the reason a long working session doesn’t cost what a naive reading of the pricing page would predict.

Two habits do the bulk of the work, and neither needs a spreadsheet. Open the narrowest folder that actually contains the task, rather than pointing Claude at a shared drive and letting it wade. And start a fresh conversation when the subject changes, instead of letting one session carry an afternoon of unrelated history — see sessions for when to clear and when to keep going. When you want the real number rather than a feeling, /cost tells you what the current session has used.

One question that gets filed under cost but isn’t really: whether you need a more powerful computer. For almost every professional the answer is no — the model runs in a data centre and your laptop’s job is to display the result. The guide below covers the narrow cases where local hardware genuinely changes the answer, and why the current hardware market makes the cloud case stronger rather than weaker.

The words

Prompt caching Context caching: Prompt caching saves the model's processed version of text you send repeatedly — a long set of instructions, a big document — so it doesn't have to re-read it from scratch every turn. The next request reuses the cached part, which is much faster and far cheaper, since cached input is billed at a fraction of normal input. In a long Claude Code session, where the same context is reused turn after turn, it's one of the biggest cost savers there is.

How to do it

/cost Show how many tokens this session has used, and the cost so far.

The words

Read more

How to do it