Claude's paid plans run on two meters: a session limit that resets every five hours and a weekly limit, with a separate weekly allowance for Fable. What drains them is what Claude has to read and how hard you ask it to think: long threads, PDFs sent as page images, high effort, tools left on. The biggest single fix for document work is converting PDFs and decks to Markdown before you upload them.
It's 2pm on a Tuesday. The message box goes grey, and the report you were halfway through is stuck until 5.
Most of our clients who moved to Claude in the past six months now get asked for extra usage almost daily. IT or HR ends up managing the frustration and the overspend on top of the adoption work. After all the effort to get a team using its licences, that's a frustrating place to land.
Buying more usage keeps things moving. My first question is what we're asking Claude to read.
I first wrote this list in April 2026. Since then Anthropic shipped Opus 5.5, three Fable releases, a weekly Fable limit, usage credits and a usage screen in Claude Code that names what's eating your plan. This version, updated on 12 November 2026, follows the current help pages. Every number has a link, or a label saying it's an assumption.
The short version
- Two limits, four dials. A five-hour session limit and a weekly limit. Conversation length, features, model and effort level decide how fast you reach them (Anthropic's help page).
- The file is often heavier than the prompt. Claude reads a PDF as extracted text plus an image of every page. In a worked example, a 30-slide deck weighs 60,000 tokens as a PDF and 15,000 as Markdown. Those are assumptions, shown below.
- Convert before you upload. Microsoft's MarkItDown turns PDF, PowerPoint, Word and Excel into Markdown on your own machine, in two commands. I built a small local app on top of it, Plainly, for people who don't open a terminal.
- Opus 5.5 by default, Fable by hand. Fable drains limits faster than any other Claude model, and Anthropic says to start most work on Opus 5.5.
- In Claude Code, ask the tool.
/usagelists what's behind your consumption and offers to fix part of it.
How Claude's usage limits work in November 2026
Paid plans (Pro, Max, Ultra, Team and seat-based Enterprise) share one allowance across Claude on the web, Claude Desktop, Cowork and Claude Code. The session limit resets five hours after your first message. The weekly limit resets at a fixed time for your account, and some plans carry a second weekly limit for Fable alone (Claude Help Center, how usage and length limits work).
Anthropic lists four things that move the meter: how long and complex the conversation is, which features you use, which model you pick, and the effort level. There's no published exchange rate between tokens and the percentage bar. Both bars are under Settings, then Usage.
When you hit a limit you wait for the reset, or you turn on usage credits and keep working at a metered rate. Most of the "can we pay for more?" requests come from there.
What changed since the April version
| Topic | April 2026 | November 2026 |
|---|---|---|
| Default model | Haiku for routine work, Sonnet and Opus above | Opus 5.5 is Anthropic's "strongest daily driver". Sonnet 5.5 and Haiku 5.5 sit under it, Fable 5.2 above. |
| Top model | Opus | Fable 5.2, with its own weekly limit. On Pro and standard Team seats it runs on usage credits from 16 November 2026. |
| Thinking | A toggle you switch off | Always on for the 5.5 models and Fable. The dial is the effort level. |
| Past the limit | Wait | Wait, or pay with usage credits, capped by a spend limit. |
| Long chats | Hit the length limit | Claude summarises earlier messages and carries on. It costs usage each time. |
| Claude Code | /compact early and often | /usage shows the breakdown. /clear is free, /compact is a large request. |
Model positioning and the Fable rules are in our Claude Fable 5.2 business guide, with Anthropic's sources. The rest comes from the help pages linked at the bottom.
Start with the file: 4 tactics for documents
1. Convert PDFs and decks to Markdown before you upload
"Summarise this report" is a short request. Claude still has to process the report.
When Claude reads a PDF visually, each page becomes an image, and the text extracted from that page is sent alongside it. Anthropic's documentation counts 1,500 to 3,000 text tokens per page depending on density, plus the image cost of the page (Claude Platform docs, PDF support). Useful for a chart. Extra input when all you need is the wording.
A PowerPoint exported to PDF carries the same overhead, slide after slide. Markdown is plain text with simple markers for headings and lists. The file ends in .md and keeps the words and their structure without a picture of every page.
| File | What Claude receives | Tokens (worked example) |
|---|---|---|
| 5-page report, Markdown or clean Word | Text | 4,100 |
| Same report, PDF read visually | Text plus 5 page images | 11,600 |
| 30-slide deck, Markdown | Text | 15,000 |
| Same deck, exported to PDF | Text plus 30 slide images | 60,000 |
These are illustrative assumptions, not measurements from client files: 4,100 text tokens for the report, 15,000 for the deck, and 1,500 visual tokens per rendered page or slide. For the deck, 15,000 + (30 × 1,500) = 60,000. Real image costs depend on the model and the resolution (Claude Platform docs, vision).
In that example, dropping the slide images removes 45,000 input tokens. It doesn't promise four times your allowance, because the file is only part of what Claude processes. The full reasoning is in Stop hitting Claude AI usage limits.
Three ways to turn a PDF, PPTX or DOCX into Markdown
Route 1: two commands, on your own machine. MarkItDown is Microsoft's free, open-source converter (MIT licence, Python 3.10 or later). It handles PDF, PowerPoint, Word, Excel, HTML, CSV and more. Install it once:
pip install 'markitdown[all]'
Then convert a file:
markitdown quarterly-review.pptx -o quarterly-review.md
Nothing leaves the computer. The .md file lands next to the original, ready to check and upload.
Route 2: Plainly, for people who don't open a terminal. I built Plainly for my own client work. It opens in a browser, and the conversion runs on my Mac. I drop in a document, check the extracted text and download the Markdown. It's MarkItDown with an interface around it, and the fuller version also reads scanned pages locally.
You can build the starter version yourself in Claude Code:
- Create an empty folder called Plainly.
- In Claude Desktop, open the Code tab, choose Local and select that folder.
- Choose Manual permissions, so you review each change as Claude makes it (Claude Code docs, desktop).
- Paste the build prompt from the Plainly article.
Building it uses your Claude allowance once. Running the finished converter doesn't call an AI at all.
Route 3: a whole folder, when Claude Code or Cowork is already installed. Ask for the conversion as a job and keep Claude out of the files:
Install Microsoft MarkItDown if it's missing. Convert every PDF, PPTX and DOCX in this folder to Markdown, saved next to the original with the same name. Run the converter from the command line. Don't open or read the documents yourself. List any file that failed.
The last two sentences matter. Without them Claude may read each document to "check" it, which is the cost you were trying to avoid.
What about online converters? They exist and many are free. Some upload your document to another provider before the result ever reaches Claude. For a public report that may be fine. For a salary grid or a client contract, I want to know who gets the file and what they keep.
Uploading the finished Markdown to Claude still sends its contents to Claude. Your company's rules about what belongs there still apply.
2. Send the pages the job needs
For rewriting slide copy, the extracted text is enough. For a design review, Claude needs the slides. Sending the six slides under discussion keeps the other 24 out of the request.
The same goes for a 90-page contract when the question is about clause 14.
3. Check the conversion before you rely on it
Tables can lose their reading order. A chart's labels can survive while the relationship between them disappears. Scanned pages come out empty unless the converter runs OCR.
I keep the original nearby and add back the page that carries what the text has lost. A smaller file with a broken table creates another job.
A clean Word document is already a sensible choice for text work. Markdown earns its place when the conversion removes visual material the task doesn't need.
4. Put recurring documents in a Project
The brand guide uploaded into five chats is processed five times. In a Project it's uploaded once and cached: reused content counts less against your limits while the cache is warm. Caches expire after a period of inactivity, so the first message after a long break counts the content in full again (Claude Help Center, usage limit best practices).
Larger Projects switch to retrieval and load only the relevant passages. Short project instructions and a tidy file list help: Anthropic's own advice is to remove project files nobody uses. Upload the Markdown versions and you get both savings.
Conversation habits: 5 tactics
5. One job, one conversation
Follow-ups carry the earlier context, attached report included. Caching softens the cost of the repeated part, so counting every follow-up as a fresh full-price upload would overstate it. The thread still grows with every turn.
When a chat nears the context limit, Claude now summarises earlier messages and carries on. Anthropic notes that long conversations that trigger this consume more of your usage limit, and that starting a new chat can help.
When the job changes, I open a new chat. If I need continuity I ask for a ten-line summary and paste it in. On paid plans, memory and chat search can also bring back what was said elsewhere.
6. Batch related questions
Market size, top three competitors, their pricing: one message, one table. Three separate messages re-send the thread three times. Batching is the first tip on Anthropic's best-practices page.
7. Edit the message. Don't send a correction
"Actually, I meant..." keeps the wrong prompt and the wrong answer in the thread for good. Editing your message replaces that turn. Our team rule: edit for fixes, new message for new steps.
8. Say how long the answer should be
"Summarise this report" can come back at 1,500 words when you needed 200. The extra words are output, and they stay in the thread for every later turn.
"Eight bullets, 200 words maximum, sections 2 and 4 only" costs one line to type.
9. Outline first, then expand
A wrong first draft costs the draft plus the rewrite. Asking for the outline, correcting it, then expanding section by section moves the correction to the cheapest moment. "Make it better" does the opposite: it buys another full pass with no criteria.
Model, effort and tools: 4 tactics
10. Keep Opus 5.5 as the default. Pick Fable by hand
Fable models consume plan limits faster than other Claude models, and Anthropic's launch post says to start most work on Opus 5.5. On Max, Ultra and premium Team seats, Fable is capped at half the weekly limit. On Pro and standard Team seats it runs on usage credits from 16 November 2026 (Claude Help Center, Fable 5.2 launch offers).
Our rule, and it's an opinion: Opus 5.5 for daily work, Sonnet 5.5 or Haiku 5.5 for emails, summaries and formatting, Fable 5.2 for the job that failed twice on Opus. The table of which job goes where is in the Fable 5.2 guide.
11. Lower the effort level before you blame the plan
On Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable, thinking can't be switched off. The effort level decides how much of it you pay for, and thinking tokens count as output (Claude Code docs, manage costs).
I raise effort for one hard job, then put it back. An email draft at maximum effort is a helicopter to the grocery store.
12. Turn off web search and connectors you aren't using
Anthropic's help page flags tools as token-intensive and suggests turning off connected apps and web search when you don't need them. A brainstorm doesn't need your CRM attached.
13. Give research its own conversation
Research runs many searches and reads many pages. I run it in a dedicated chat, export the findings, and write from them in a clean one with the tools off.
Claude Code and Cowork: 4 tactics
14. Run /usage and read the recommendations
On a paid plan, /usage in Claude Code now shows what counts against your limits: skills, subagents, plugins, each MCP server, with a share for each. Under that sits a list of recommendations, such as trimming a CLAUDE.md that adds 49,000 tokens to every request. Press f and Claude Code applies the ones it can (version 2.1.305 or later).
The tool tells you where the usage went.
15. /clear between tasks. /compact with care
Stale context is re-sent with every message. /clear starts fresh and costs nothing. /compact has to read the whole conversation to summarise it, so compacting a large context is itself a large request.
The April version of this article said to compact at 50% of the context. Anthropic's current guidance is simpler: clear when the task changes, compact when you need the continuity.
16. Keep CLAUDE.md under 200 lines
CLAUDE.md loads at the start of every session, relevant or not. Anthropic's advice is to keep it under 200 lines and move workflow-specific instructions into skills, which load only when they're used.
17. Cap the subagents
Every subagent sends its own requests on top of the main conversation. Agent teams use roughly seven times the tokens of a standard session when teammates run in plan mode, per Anthropic's documentation. I write the cap into the brief: "use at most four subagents".
And plan before the work starts. A correction after an hour of work costs an hour of tokens.
For the whole team: 1 tactic
18. Set the defaults once, in the admin console
Habits spread slowly. Settings apply on Monday.
- Model defaults. Admins choose which model new conversations start on, across chat, Cowork and Claude Code.
- Usage credits with a ceiling. Spend limits exist at organisation, group and member level. People keep working past the plan limit and finance keeps a cap.
- The spend report. Estimated spend per user and per model, with a CSV export. It shows who needs a premium seat and who needs a converter.
- One page of rules. Convert before upload. One job per chat. Opus by default. Effort back to medium.
Where each of these lives, screen by screen, is in our guide to AI spend dashboards. If the question is which seat to buy for whom, see ChatGPT vs Claude plans for companies.
If you run Claude for a team
The settings cap the bill. They don't change what people ask Claude to read. When the same people hit the limit every week, look at the work before the plan: files uploaded raw instead of converted, one long chat carrying five jobs, effort left on high for routine work. That's what we look at first, team by team, in our Claude rollouts for teams, and we measure usage again at 30 days.
What I took out of the April version
The first version said teams waste 50 to 70% of their token budget and get two to three times more work from the same plan. It also put multipliers on thinking, search and connectors. I couldn't trace any of those figures to a source, so they're gone.
Anthropic publishes no token-to-percentage rate, and I haven't measured a before and after on a client team. What I can say: the tactics above are the ones Anthropic's own pages describe, and the document one is the first I check when a team tells me the plan is too small.
The plan is rarely my first suspect. Nobody showed them what a PDF weighs.
Frequently Asked Questions
Why do I hit Claude's usage limits so quickly?
Four things drive usage, according to Anthropic: conversation length and complexity, the features you use, the model, and the effort level. In practice the usual causes are one long thread used for everything, PDFs and decks sent as page images, Fable or high effort used for routine work, and web search or connectors left on.
How do Claude's usage limits work in 2026?
Paid plans have a session limit that resets five hours after your first message, and a weekly limit that resets at a fixed time. Some plans add a separate weekly limit for Fable. Usage is shared across Claude on the web, Claude Desktop, Cowork and Claude Code. You can see both bars under Settings, then Usage.
How do I convert a PDF or PowerPoint to Markdown for Claude?
Install Microsoft's open-source MarkItDown with pip install 'markitdown[all]', then run markitdown file.pptx -o file.md. It converts PDF, PowerPoint, Word and Excel on your own machine. If you'd rather not use a terminal, you can build a small local converter such as Plainly in Claude Code, or ask Claude Code or Cowork to convert a whole folder from the command line.
Does converting a PDF to Markdown lose information?
It can. Tables may lose their reading order, charts keep their labels and lose the relationships, and scanned pages need OCR. Keep the original and add back the pages that carry visual information the task needs.
Does editing my last message save usage compared with sending a correction?
Yes. Editing replaces the turn, so the mistaken prompt and its answer leave the conversation. A follow-up correction keeps both in the thread, and they're re-sent with every later message.
What are Claude usage credits?
Usage credits let people on Pro, Max, Ultra, Team and seat-based Enterprise plans keep working after they reach a plan limit, at a metered rate. Admins can set spend limits for the organisation, a group or one member. From 16 November 2026, Fable 5.2 on Pro and standard Team seats runs on usage credits.
Can a team get more out of Claude without upgrading the plan?
Often, yes. Converting documents to Markdown, keeping one job per conversation, defaulting to Opus 5.5 at medium effort and switching off unused tools all reduce what Claude has to process. Anthropic doesn't publish a conversion rate, so measure with the usage bars before and after.
Sources
- Claude Help Center, How do usage and length limits work?
- Claude Help Center, Usage limit best practices
- Claude Help Center, Claude Fable 5.2 launch offers
- Claude Platform docs, PDF support and Vision
- Claude Code docs, Manage costs effectively and Claude Code on desktop
- Microsoft, MarkItDown on GitHub
- Toni Dos Santos, Stop hitting Claude AI usage limits, Vibe Working, 28 October 2026
Free, 8 minutes
Before you buy more usage, find out what's holding usage back.
The AI Maturity Audit places your company on the WCS AI Maturity Model in 21 questions.
You get your stage, the barrier in the way (people, leadership, process or foundations) and what to fix first.
Free personalised report, no credit card.
Run the free AI Maturity Audit →