As per today's date (27 September 2026) — AI prices and plan limits change often, so every price below links to the official page. Check it before you decide.
I'm on Claude's Pro plan, and I haven't hit the session limit once in the last three months. That isn't because I use it lightly — it's because of a handful of habits that all come from one thing: understanding how usage is actually counted.
Most people treat an AI subscription like Netflix: pay, use, never think about it. But AI isn't billed by the hour or by the number of questions. It's billed by tokens — small pieces of words — and almost everything you do in a chat changes how many you burn. Here are the seven things I wish I'd known before paying, and the habits that keep me under my limits.
1. You pay for the answer more than the question
On the pay-as-you-go API, output (what the AI writes) costs five times as much as input (what you send). Claude's official pricing as of today:
| Model | Input / 1M tokens | Output / 1M tokens | Good for |
|---|---|---|---|
| Claude Haiku 4.5 | $1 | $5 | Quick fixes, reformatting, simple questions |
| Claude Sonnet 5 | $2 | $10 | Most everyday work and coding |
| Claude Opus 5.5 | $4 | $20 | Hard reasoning, complex code, long analysis |
On a monthly plan you don't see these numbers, but long answers still eat your allowance faster. My habit: I say how long I want the answer — "one line", "3 bullets", "just the code". Most of the time that's all I needed anyway.
2. Every message re-reads the whole chat
This is the one almost nobody knows. The AI doesn't remember your conversation between messages; every time you hit send, the entire chat so far is processed again. So message 30 in a long chat costs far more than message 1, even if both are one short question:
Earlier chat the AI re-readsThis message (question + answer)
| Message # | Earlier chat re-read | New question + answer | Total processed |
|---|---|---|---|
| 1 | 0 | 500 | 500 |
| 5 | 2,000 | 500 | 2,500 |
| 10 | 4,500 | 500 | 5,000 |
| 20 | 9,500 | 500 | 10,000 |
| 30 | 14,500 | 500 | 15,000 |
Across the whole 30-message chat that's 232,500 tokens processed — but only 15,000 of them are new questions and answers. The other 94% is re-reading. Anthropic says the same thing in plainer words: your limit depends on the length of your current conversation.
My habit: when a chat drifts to a new topic, I start a fresh one and paste a two-line summary of what matters.
A recent example: I asked Claude to analyse a whole project — every file. Instead of doing my coding tasks in that same heavy chat, I asked it for a summary of the points that mattered, started a new chat with just that summary, and did the coding task there. Then I did the same for each task after it. Each new chat starts light instead of dragging the entire project along with every message.
3. Files, screenshots and pasted text ride along every turn
Whatever you attach becomes part of the chat, so it gets re-read with every later message too. Images are converted to tokens by size. From Anthropic's vision documentation:
| Image size | Tokens (approx.) |
|---|---|
| 200×200 px (a cropped button or error) | 64 |
| 1000×1000 px | 1,296 |
| 1920×1080 px (a full-screen screenshot, newest models) | 2,691 |
My habits: I crop screenshots to just the error with our image cropper, paste only the relevant section of a document (our PDF to text tool makes that easy), and check how long something is with the word counter before pasting it. For documents I reuse, I keep them in a Project — Anthropic notes that cached project content counts less against your limits than new content.
4. The most powerful model isn't always the right one
Look at the table in point 1 again: the top model costs four times the smallest one per token. Fixing grammar, renaming variables or turning notes into bullets doesn't need it.
My rule: I use Sonnet for straightforward tasks — when the bug is clear, I understand the fix, and I can tell the AI exactly what to do. If I get stuck, or realise the problem is more complex than it looked, I switch to Opus. Most of my work never needs that switch.
5. Extras quietly drain your limit
Claude's help center lists what affects usage limits: message length, attachment size, conversation length, tool use like Research and web search, model choice and effort level. Web search, deep research and high "thinking" or effort settings are brilliant for hard questions and wasteful for easy ones.
My habit: I leave web search and extended thinking off by default and switch them on only when the question actually needs them.
6. Fix the prompt instead of piling on corrections
Every "no, I meant…" adds another turn that the whole chat gets re-read for (see point 2). A clear first message with context, the format you want and the length you want is cheaper than three corrections. If the answer misses, I edit my original message rather than replying underneath it, and I don't hammer the regenerate button — each regeneration is a full new answer.
7. Watch the meter, and pick the plan for how you actually work
As of today, Claude's plans are Free, Pro at $20/month ($17/month billed yearly) and Max from $100/month for 5× or 20× Pro's usage. The usage meter tells you whether you really need to upgrade or just change habits first.
For me the answer was clear: with the habits above, the Pro plan is enough, and I haven't needed Max.
If you're a developer paying for the API directly, two more savings: batch processing costs 50% less for work that doesn't need an instant reply, and prompt caching makes repeated input much cheaper.
The short version
- Ask for short answers.
- Start a new chat when the topic changes.
- Paste and attach only what's needed — crop screenshots.
- Use the smallest model that does the job.
- Turn on web search, research and deep thinking only when needed.
- Edit your prompt instead of stacking corrections.
- Check your usage meter before upgrading.