> 📥 The full guide is available as a PDF (Danish, 8 pages). Download it at the top of the page — all prices, limits and sources in one place.
Last month I gave Claude Fable 5 one big job: a complete workshop package for a leadership team. Agenda, talking points, exercises, follow-up. One prompt, one answer.
The result was the best I've ever gotten out of an AI model. And when I opened my usage overview afterwards, I could see why. That single task had taken a solid bite out of the week's quota. I had to top up with more credits.
That's the new reality with Claude: the smartest model on the market now sits behind a taximeter. Used right, it's worth every krone. Used as your everyday chat model, you may find yourself feeding the Claude meter.
This guide is about choosing right: model, effort, plan and habits. Those are the four knobs that decide whether Claude is a good investment or a hole in the budget.
1. Four models: the partner shouldn't write the minutes
Claude comes in four levels of "intelligence". Think of it as a consulting firm.
- Haiku 4.5 — the student assistant. Fast and cheap. Use it for the trivial.
- Sonnet 5 — the capable project manager. Balance of speed and quality. Use it for most of everyday work.
- Opus 4.8 — the senior consultant on retainer. Very smart and included in the subscription, but with its own weekly cap, separate from the others.
- Fable 5 — the external partner on an hourly rate. The smartest right now, and typically costs extra when you push it.
The partner sets strategy and stress-tests the big decisions. The minutes, the email and the small edits belong with the others.
(Anthropic also has a model called Mythos 5. It's the same underlying model as Fable 5, reserved for specially approved organizations. For the rest of us, the top is Fable 5.)
The big mistake right now? People discover how good Fable 5 is, set it as their default model and chat away. It's the most expensive habit you can pick up in 2026 — also environmentally, in energy and water. You don't hire a partner at 3,000 kroner an hour to fix commas either.
2. The bill grows with the thread, not with the question
Fable 5 is billed in tokens: 10 dollars per million input tokens and 50 dollars per million output tokens. A token is roughly three-quarters of an English word. The crucial part is that Claude re-reads the entire conversation every time you send a new message. Your turn number 20 pays for the previous 19 all over again.
> Napkin math, no caching, round numbers, Fable 5. About 7 cents for one question with a thorough 700-word answer. 5-6 dollars for a working session of 20 turns. Around 20 dollars for a marathon thread of 40 turns — for a single conversation. Short threads are cheap, long threads are expensive. New task, new chat.
And a detail few people know: the newest models (Fable 5, Sonnet 5 and Opus from version 4.7) use a new tokenizer that produces roughly 30 percent more tokens for the same text. Same paragraph, more tokens, bigger draw on the quota.
The language tax. Danish takes up so little space in the models' training data that the tokenizer chops Danish words into more pieces than English ones. An analysis of OpenAI's newest tokenizer measured 31-43 percent more tokens for languages like French, German and Italian; Danish wasn't in the measurement, but almost certainly sits in the same range — around 40 percent on top of English. The direction is certain, the precise number is an estimate.
> Tip. Ask Claude to summarize long conversations into a new thread. It's happy to, and you start fresh on a shorter thread: you save tokens and money — and spare the environment (less energy and water use).
3. Your plan decides whether Fable is included at all
After a summer in which Anthropic moved deadlines several times, the line is now fixed. Status as of July 21, 2026:
If you have Max, a Premium seat on Team or a Premium seat on Enterprise, Fable 5 is included in the subscription. You can use it for up to 50 percent of your weekly limits at no extra cost. And note: Fable draws on the same limits as the rest, just faster than the other models.
If you have Pro or a Standard seat on Team, Fable 5 is not part of your quota. You pay per use via usage credits, at the same rate as the API: 10 and 50 dollars per million tokens. A consolation: Pro and Standard users get a one-time 100-dollar credit, confirmed by Anthropic on July 20.
Today's prices, in USD excl. VAT: Pro costs 17 dollars a month billed annually, 20 monthly. Max starts at 100 dollars for 5x Pro usage and comes in a 20x version. Team costs 20 dollars per seat per month annually for Standard seats and 100 for Premium seats.
And then the limits, which surprise many: your usage is counted in rolling 5-hour windows, with two weekly caps on top: one for Opus alone and one for all other models. So Opus doesn't cost extra, but it has its own weekly quota and can also run dry. Everything draws from the same pool: chat on the web, the desktop app, mobile, Claude Code and Cowork. That's why a wild Cowork afternoon can shut down your evening chat.
My own setup is a Team plan for TryZone as the primary account, with one Premium and one Standard seat alongside. Team gives the data protection and admin control (no model training by default), Premium gives the quota for heavy daily work. For serious work use, that combination is hard to beat.
4. Effort is the knob most people never touch
Beyond model choice, Claude has an effort setting. In the API the levels are low, medium, high, xhigh and max; in the app there's a corresponding selector next to the model picker. It controls how long the model thinks before answering. More thinking time gives better answers and a bigger draw on the quota. The default is high, and that fits most things.
Dial down to medium or low when the task is routine: rewrites, summaries, small edits. The answers are still good, they arrive faster, and they cost less.
Dial up to xhigh or max when the task is genuinely hard: complex analysis, long agent tasks, code that has to work first time.
> My rule of thumb. Fable 5 on low effort often beats older models at full power. Low effort on a smart model is frequently a better buy than full effort on a cheaper one.
5. Give Fable goals, keep it to few turns, then switch
The most important workflow in this entire guide has three steps.
1. Start in Fable 5 with one big, well-defined goal. All context in the first message: files, requirements, deadline, audience. Ask it to question you before it starts.
2. Let it plan and deliver. One to two turns, maybe three.
3. Switch model to Opus 4.8 in the same conversation, and run execution and polishing there. Opus is included in the subscription, reads the whole thread and follows the plan faithfully.
Old models had to be fed small tasks, one at a time. The new ones need goals.
A task: "Write a follow-up email to the client."
A goal:
```
The leadership team must choose three AI initiatives at a two-hour meeting on
August 14. Here's the context and the participant list. Plan the meeting, build
the material, and ask me for anything you're missing before you start.
```
Fable rewards goals with plans, pushback and work it has checked itself. It asks before it builds. That's what you're paying the partner rate for.
One caveat: in Cowork you currently can't switch models mid-session, so there you have to choose right from the start — or use the trick above and ask Fable to pack a summary into a fresh thread for Opus.
6. Three tasks where Fable is worth the money
1. The decision that's expensive to reverse. Give it the decision, your priorities in order, and the constraints you can't change. Ask it to interview you until it understands the decision better than you formulated it. Then demand a recommendation, the strongest argument for the rejected option, and the most likely point at which the recommendation turns out to be wrong. Board-level sparring, typically for under a dollar.
2. The research you have to act on. Ask for a report that opens with the answer, require sources for every material claim, and have it separate evidence from inference. Finally, ask it to attack its three load-bearing claims as a skeptic before delivering. That self-check was pure fantasy on older models. I use a fixed QA skill for exactly that work.
3. Messy material that has to become a plan. The week's meeting notes, email threads and half-finished documents in; one plan out, with commitments, decisions and conflicts, all traceable back to the source. That kind of synthesis is Fable's home turf.
Common to all three: you couldn't have given the task to an intern. That's the test. If an intern could have done it, it's a Sonnet task.
7. Nine habits that keep the bill down
1. New chat per task. Never reuse a long thread for a new topic.
2. All context in the first message. Ten clarifying turns cost more than one thorough briefing.
3. Give goals, not tasks. Big questions to the expensive model, small ones to the cheap ones.
4. Switch models along the way. Fable plans, Opus executes, Sonnet tidies up.
5. Use effort actively. Down for routine, up for the hard stuff.
6. Put fixed context in Projects or skills, so you don't pay to repeat yourself in every chat.
7. Check Settings > Usage often. You can't manage what you don't measure.
8. Set a cap on your usage credits before you experiment with Fable. Decide the amount while you're sober.
9. If you have a team: set spend limits per user in the organization settings before someone discovers how fun it is to give Fable marathon tasks.
8. How I run it myself
Fable 5 gets two kinds of tasks from me: decisions that need stress-testing, and plans that must be built from scratch. Both in one to two turns, always with all context in the first message.
Opus 4.8 does all the long work: documents, decks, analyses. It's my workhorse, and it's included in the subscription. Sonnet 5 takes the quick questions. I rarely touch Haiku myself; it's most interesting if you build automations on the API.
The result: Fable quality on the decisions where it matters, without the weekly quota disappearing into small stuff.
9. Three questions before the next bill
1. If you pay yourself: Do you know how much of your usage goes to conversations a cheaper model could have handled?
2. If you have a team: Who has access to Fable 5 today, and what do they use it for?
3. And in both cases: When did you last open the Usage page?
The answers take ten minutes to find. They're the best-paid ten minutes in your AI budget this month.
Sources
All prices and figures are based on the following sources (USD, excl. VAT, status July 2026):
- claude.com/pricing (plans, API pricing and usage FAQ, retrieved July 21, 2026)
- support.claude.com · Claude Fable 5 on your plan (article 15424964) · Usage limit best practices (article 9797557)
- support.claude.com · Manage usage credits for paid Claude plans (article 12429409)
- platform.claude.com/docs · Pricing (model pricing, tokenizer note, caching and batch)
- platform.claude.com/docs · Effort (effort levels and recommendations)
- anthropic.com/news/claude-fable-5-mythos-5 (the difference between Fable 5 and Mythos 5)
- x.com/claudeai · official statement: Fable 5 included in Max and Team Premium, one-time credit for Pro/Standard (July 20, 2026)
- kathane.substack.com · Not speaking English to ChatGPT costs you millions of tokens (o200k tokenizer analysis, November 2025)
- arxiv.org/abs/2305.15425 · Petrov et al., Language Model Tokenizers Introduce Unfairness Between Languages (NeurIPS 2023)
> ⚠️ Disclaimer. All prices are in USD excl. VAT and may have changed; always check the vendors' own pricing pages. Plans, limits and model access change continuously, so verify against Anthropic's own pages before making decisions on these numbers. General practical guidance, not legal advice. Status as of July 2026.
📬 Subscribe to the newsletter "AI, Built Human" on Substack — weekly insights on AI in practice.
Stefano Vincenti · GenAI strategist and architect · External lecturer, IT University of Copenhagen & DIS Copenhagen · Cofounder & CTO BotTellMe · Partner, TryZone · aitrainer.dk