The classic dilemma: a smaller model thinking harder or a bigger model thinking less? Same prompt, same tools, same machine. Green = the one that used less quota.
Simple: one clear goal, a file or a page or two, an answer you can check, done in minutes. Ten examples - the three marked tested are the ones measured on this page; the others are jobs of the same size, shown for scale - they were not run, so their costs are not measured.
Write parse_duration("1h30m") that returns seconds and rejects bad input
Add a "copy to clipboard" button to one existing HTML page
Fix a failing unit test whose error message points to the bug
Rename a function and update every call to it in a small repo
Find the latest SQLite, Python and PostgreSQL versions and their release dates
Look up the current Claude Pro and Max plan prices on the official pricing page
Find which Node.js version is the current LTS and when its support ends
Save every book in the Mystery category of books.toscrape.com (2 pages) to a CSV
Pull the Hacker News front-page headlines and links into a JSON file
Get a product's name, price and stock status from one shop page
Medium: several files, sources or pages, a spec to follow, many small facts that all have to be right. Again the three marked tested were measured; the rest are shown for scale.
Make a 6-file shop app follow its README: fix 2 bugs (negative stock, VAT rounding), add discount codes and a CSV export command
Add pagination to an existing API endpoint and update its tests
Write unit tests for an untested 200-line module and fix what they uncover
Move a small script from one HTTP library to another without breaking its tests
Find the highest stable version and release date of 8 runtimes: Python, Node.js, Go, Rust, PHP, Ruby, PostgreSQL, SQLite
Compare the cheapest plan of 5 VPS providers: price, RAM, CPU and traffic
Summarize the breaking changes between two major versions of a library from its changelog
Scrape all 65 Fiction books on books.toscrape.com over 4 pages, opening every detail page for UPC, rating and stock
Collect a paginated job board into one CSV and remove the duplicates
Pull product name, price and stock from 3 different shop sites into one table
Each cell is one real run in Claude Code. The big number is what the run would cost at Anthropic's API list prices - the best public proxy for how much plan quota it eats (see how quota works). × cheapest compares it to the cheapest run that got a perfect score on the same task; redder = more quota. The badge is the graded score. Click a cell for the actual output.
Five stars = perfect result for at most 1.5× the quota of the cheapest perfect run. Each further step (3×, 6×, 12×) costs a star. A run that got anything wrong gets 2 stars if it was at least 80% right, otherwise 1 - a cheap wrong answer is not good value.
The same 9 setups (Haiku, Sonnet and Opus at low, medium and high) ran both task sets. Quota = all three tasks added up at API list prices. Quota growth = medium cost divided by simple cost. Fable and xhigh/max were left out of the medium run.
What the 87 runs above mean for a normal working day with easy, medium and hard jobs mixed together. Rows marked not measured are advice for jobs this page did not test.
| The job | Start with | Step up to | Why (measured) |
|---|---|---|---|
| Small function, one-file edit, quick lookup | Sonnet · low / medium | Sonnet · high | Perfect on all simple tasks for $0.21; higher effort only added cost |
| Multi-file feature or bug fix with a clear spec | Sonnet · medium | Opus · low / medium | Both 28/28 on the medium coding task; Sonnet $0.10-0.12, Opus $0.12-0.20 |
| Research with many facts to get right | Opus · low | Opus · medium / high | Opus 16/16 every time in 11-12 turns ($0.29-0.36); Sonnet also 16/16 but swung up to $0.62 |
| Scraping, bulk data, repetitive edits | Sonnet · low + ask for a script | Opus · low | 65 books with detail pages for $0.04: the script does the work, not the model |
| Hard, vague or architectural problems | Opus · high in plan mode | Opus · xhigh, then Fable | not measured Hard tasks were not tested; this is where the bigger setups should earn their price |
| Avoid as a default | max effort · Haiku for research | – | max cost 4-5× high with no extra points; Haiku research scored 2/16 twice |
Move up one rung only after a wrong or weak answer. On every task measured here the lower rungs already scored 100%, so the higher rungs were paid for and not needed.
/clear between unrelated tasksEvery turn re-reads the whole conversation. A fresh context keeps each turn small. Use /compact to keep going on the same task with a shorter history.
/model and /effort change the setup. Prompt caches belong to one model, so switching in the middle of a long task means the history is processed again. Switch right after /clear.
The opusplan model setting uses Opus in plan mode and Sonnet when it executes. Subagents can run a cheaper model with model: sonnet in their frontmatter.
For data jobs say "write a script that fetches and parses". Sonnet and Opus did that and scraped 65 books for 4-7 cents; Haiku on high kept going for 39 turns and 28 cents.
Every prompt here named the file, the columns or the line format, so a run could finish without asking back. Each extra turn re-reads the whole context.
/usage and /context/usage shows your plan limits and what counts against them; /context shows what fills the context window. Look before a long session, not after the limit hits.
| Task | Playbook pick | Quota | Opus · high for everything | Haiku for everything |
|---|---|---|---|---|
| Small function (simple coding) | Sonnet · low | $0.033 | $0.089 | $0.046 · 20/20 |
| Fix + extend the shop app (medium coding) | Sonnet · medium | $0.118 | $0.198 | $0.153 · 28/28 |
| Look up 8 runtimes (medium research) | Opus · low | $0.287 | $0.311 | $0.477 · 2/16 |
| Scrape 65 books (medium scraping) | Sonnet · low | $0.039 | $0.074 | $0.275 · 65/65 |
| Day total | all correct | $0.48 | $0.67 · all correct | $0.95 · research wrong |
Haiku runs are at effort high, the setting that gave Haiku its best scores. Picking per task used 29% less quota than Opus on high for everything, and about half of "always Haiku", which also got the research wrong.
What Anthropic says officially, and the price list we used to turn tokens into quota.
Anthropic's help center: usage depends on "which Claude model you're chatting with, and the effort level you've selected" - and "Higher effort uses more tokens." Source
Pro and Max have a five-hour session limit and a weekly limit, shared between claude.ai, Claude Code and the IDE. Max 5x and 20x give 5 and 20 times Pro's per-session allowance. Source
Fable models "draw from your plan's regular weekly usage limits and use them faster than other Claude models"; on some plans up to 50% of the weekly limit can go to Fable. Source
Anthropic does not publish how many tokens of which model make up 1% of your quota. That is why this page uses API list-price cost as the yardstick: it weighs input, output, cache and model price the way the bill would. Treat the ratios, not the dollars, as the answer.
| Model | Input $/1M tokens | Output $/1M tokens | Output price vs Sonnet | Effort levels |
|---|