# Effort level cheat sheet

Companion to *Claude Fable 5.1 in Plain English*.

---

## The five settings

`low` · `medium` · `high` · `extra high` · `maximum`

More effort means more thinking time, a higher bill and a slower answer.
It does not always mean a better answer.

---

## Which one to start on

| Kind of work | Start at | Why |
|---|---|---|
| Summarising, first drafts, rewriting tone | **low** | If it reads fine, there is no reason to pay more |
| Checking a document against rules, comparing versions | **medium** | There is a right answer and you want it found |
| Multi-step work where each step depends on the last | **medium to high** | Cutting thinking short shows up as an early wrong turn |
| Long jobs you set going and walk away from | **high or above** | Nobody is watching it go wrong |
| Anything going to a client | **your tested level, plus one** | Pennies more per document, and a wrong document is not pennies |

---

## The twenty-minute test

1. Pick one job from the last fortnight where you already know the right answer.
2. Run it on the highest setting. Keep the result as your reference.
3. Run the identical job one level down. Same wording, fresh conversation.
4. Compare the two answers **before** you look at the cost.
5. Keep dropping until it visibly breaks. Go back one. That is your floor.

Your floor is per kind of job, not per business. Re-run this the first time you
point it at genuinely different work.

---

## The numbers behind it

Measured by Cursor on their own benchmark, published 1 September 2026:

| | Score | Tokens per task | Cost per task |
|---|---|---|---|
| Fable 5.1, medium | 68.0% | 23,801 | $3.53 |
| Fable 5.1, high | 69.4% | 33,153 | $4.80 |
| Previous model, maximum | 70.5% | 103,525 | $17.32 |

Two and a half percentage points between the best and the cheapest.
Roughly four times the tokens.

Cache reads fell from $1.00 to $0.25 per million tokens on 1 September 2026.
Input ($10) and output ($50) did not change.

---

## Two questions worth asking your supplier

1. Is our usage billed per token, or on a monthly plan? The price cut only
   reaches a per-token bill directly.
2. What effort level are our automations running on, and was it tested, or is
   it just the default?

---

*Sources: Anthropic's Fable 5.1 release notes and pricing page, and Cursor's
published CursorBench 3.2 results, both retrieved 1 September 2026. Figures
change; check before quoting them.*