The effort dial: stop paying premium rates for routine work

On 24 July, Anthropic released Claude Opus 5 with something new for everyday users: a low, medium, high effort toggle. You choose how hard the model thinks on each task, and the price follows. Anthropic positions it as close to top-tier output at roughly half the cost (daily brief, 25 Jul 2026, citing Fortune and Bloomberg). The reception was split. About a third of early users praised the toggle, a quarter questioned whether the value holds at full pricing of USD 3 per million input tokens and 15 per million output (same brief, Hacker News thread).

Here is why this matters to you. You already run an effort dial in your own work. You do not give a Facebook caption the same attention as an exclusive listing pitch, and you do not prepare a routine admin email the way you prepare a course proposal. AI pricing now works the same way. If you run every task at maximum effort, you pay maximum price for work that never needed it. That is true whether you pay per token or simply burn through a subscription's usage limits faster.

How to do it: sort your own weekly tasks into three tiers before you touch the toggle. Low effort covers mechanical work on your own material: reformatting your notes, tidying your own captions, summarising an article you saved. Medium covers first drafts: your own marketing copy, a lesson outline, a market recap for your team. High is reserved for work where a wrong answer costs you money or reputation: analysis you will present under your own name, your pricing model, anything with numbers you will be quoted on. Note the pattern: everything on this list is your own workflow. Client files never go into any tier.

The lesson: the new skill is not better prompting, it is triage. Knowing which of your tasks deserve expensive thinking is the same judgment that makes a good agent or a good trainer. The worksheet below makes it a ten-minute exercise.

Artifact

Effort Dial Worksheet v1
Field artifact, Metacamp Field Notes, week of 29 Jul 2026.
Use with any AI model that offers an effort, reasoning, or model-size choice.
Scope rule: your OWN work material only. No client files, no client data, ever.

Step 1: Tier your tasks (10 minutes, once)
Write your 10 most common AI tasks into a tier. Defaults below; edit to fit.

LOW effort: mechanical work on your own material
Wrong answers are cheap and obvious. You will spot errors on sight.
· Reformat or tidy your own notes after a viewing or a class
· Rewrite your own social captions in a different length or tone
· Summarise an article or circular you saved for yourself
· Turn your own bullet list into a table, or a table into bullets
· Draft a routine calendar or admin message in your own name

MEDIUM effort: first drafts you will edit
Wrong answers cost you editing time, not money.
· First draft of your own listing marketing copy (from your own inputs)
· Outline for a lesson, workshop, or presentation you will deliver
· Weekly market recap for your own team meeting, from named sources
· Comparison table of tools or courses you are evaluating for yourself
· Draft FAQ answers for your own website

HIGH effort: judgment work published under your name
Wrong answers cost money or reputation.
· Analysis you will present publicly or be quoted on
· Your own pricing, commission, or course-fee modelling
· Long documents you must reason about end to end (policies, guides)
· Anything with calculations you will act on
· Final-pass review of a HIGH-stakes draft before it ships

Step 2: The 5 routing questions (per task, 30 seconds)
1. Will a wrong answer cost me money or reputation? Yes = HIGH.
2. Will I check every line myself before using it? No = go up one tier.
3. Is it a first draft I will rewrite anyway? Yes = MEDIUM is enough.
4. Is it mechanical (reformat, shorten, summarise)? Yes = LOW.
5. Am I unsure? Run it LOW first. Escalate only if the output fails.

Step 3: Escalation rule
Never start HIGH out of habit. Start one tier lower than you think you
need. If the output misses, move up one tier and rerun. Two failed
escalations means the problem is your input, not the effort level:
rewrite the brief, do not raise the dial.

Step 4: One-week cost log
Track for 5 working days, then re-tier.

| Day | Task | Tier used | Output OK first try? (Y/N) | Should have been tier |
|-----|------|-----------|----------------------------|-----------------------|
| Mon |      |           |                            |                       |
| Tue |      |           |                            |                       |
| Wed |      |           |                            |                       |
| Thu |      |           |                            |                       |
| Fri |      |           |                            |                       |

Read the log: if most LOW rows say Y, you were overpaying before.
If a HIGH row says N, the input was the problem. See Step 3.

Worked example (property agent, own workflow)
Task: weekly market recap for the agent's own team huddle.
Q1: wrong answer embarrassing but not costly, not HIGH.
Q2: agent checks the numbers against the source herself, stays MEDIUM.
Q3: she rewrites the opening in her own voice anyway, MEDIUM confirmed.
Result: MEDIUM effort, sources named in the prompt, 15 minutes saved,
premium effort reserved for her quarterly review deck on Friday.