One Effort Ladder, Four Models — And It Only Fits Two Of Them

The design of this cheat sheet is the clearest we have seen on this topic. Four model cards in a two-by-two grid, and running vertically between them a single coloured spine with five labels on it: Low · Medium · High · Extra · Max. The subtitle tells you how to read it: “Pick the model, then set how hard it should think.”

Two dials, four models, five settings. It is a good mental model and the five rungs are real — Anthropic’s effort documentation lists exactly five levels, and “Extra” is this card’s friendlier name for xhigh.

The trouble is that the ladder is drawn once, in the middle, as though it were the same ladder for all four. On the four models this card names, it is not.

Diagram: the same ladder against each of the four models

The effort documentation lists which models accept which level, model by model. Lining the card’s four up against those lists gives this.

The Haiku row is the one that changes how you use the card. The effort page lists the models the parameter works on, and claude-haiku-5-5 is on that list while claude-haiku-4-5 is not. On the Haiku the card names, there is no dial to turn. You get what you get.

The Sonnet row is subtler and more interesting. It has four of the five, but the missing one is “Extra”, which sits between High and Max on the card’s spine. So the ladder is not short at the top — it has a gap in the middle. Step up from High on Sonnet 4.6 and the next rung your hand finds is Max.

And the two dials are wired together

“Pick the model, then set how hard it should think” reads as two independent choices. They are not independent, because the effort dial has a different zero point on every model. From the documentation:

“Most Claude models default to high effort, spending as many tokens as needed for excellent results; Claude Opus 5.5 and Claude Haiku 5.5 default to medium.”

Which means that if you never touch the effort control — and most people never will — switching model moves it for you. The docs give the cleanest possible example of this happening inside a single model line:

“Claude Opus 5.5 supports all five effort levels, and medium is the default (Claude Opus 5 and earlier Opus models default to high, so a request that omits effort runs one level lower than it did on Claude Opus 5).”

Same company, same model family, same prompt, no setting changed by you — and one level less thinking. That is exactly the kind of thing a cheat sheet about not wasting your limits should be telling you, and a single shared spine cannot express it.

Notice what row three does to the card’s advice. The Sonnet panel says it “spends your limits the slowest”, and per token that is true. But if you arrive there from Opus, you also just moved from medium to high without deciding to. Cheaper tokens, more of them.

“2x faster” is the gap between two of the four

The caption’s warning is specific: “Pick wrong and you burn your limits 2x faster.” Two is a real number here — it is exactly the published price ratio between Opus and Sonnet. It is also the smallest gap on the card.

One important qualification, and we want to be straight about it: these are API prices, not subscription limits. Anthropic does not publish a formula converting a Pro or Max plan’s usage allowance into per-model rates, so the table above is the best public proxy for “what burns faster”, not the mechanism itself. Treat the ordering as reliable and the exact multiples as indicative.

With that said, the shape of the advice changes. “2x” sounds like a rounding error you can absorb. The distance between the top card and the bottom card is two orders of magnitude — and then effort multiplies whatever you land on, because the docs are explicit that effort “affects all tokens in the response”, thinking and tool calls included.

The joke is that the diagram is correct

Run the same check against the four models that are current today rather than the four on the card, and the picture resolves:

• Claude Fable 5.1 — all five levels
• Claude Opus 5.5 — all five levels
• Claude Sonnet 5.5 — all five levels
• Claude Haiku 5.5 — all five levels

The card’s five-rung spine is exactly right for the four models it did not name. Every name on it — Fable 5, Opus 4.8, Sonnet 4.6, Haiku 4.5 — appears in Anthropic’s legacy list, and the badge on the Fable panel reads “Free until 7th July”, a date that has been and gone. We went through what that costs you in the piece on naming a legacy model, so we will not re-argue it here. The short version is that updating the four names makes the card’s own diagram true.

It is also worth saying the card’s footer already covers this, in small grey type: “Model names, limits, free periods and defaults change often. Check Claude’s official site for what is available to you. This is a simple guide, not a benchmark.” That is an honest line, and the word doing the work in it is defaults.

Four honest caveats

1. We read availability off the documentation, not from the apps. The effort page lists supported models and which levels each accepts; a consumer app can expose fewer controls than the API does, or name them differently.
2. API prices are not plan limits. We use them because they are published and the limit formula is not. The ranking holds; treat “100×” as a scale, not a bill.
3. Lists like these move. Model line-ups changed three times this year, and a level can be added to a model without ceremony. Everything here reflects October 2026.
4. We are not claiming the card’s descriptions are wrong. “Sonnet for everyday work, Haiku for bulk, Opus for high-stakes” is sound guidance and matches the documentation’s own framing. The issue is one shared ladder drawn across four models that climb differently.

The one-line version

The five effort rungs are real, but they are not one ladder: on the four models this card names, one is missing the Extra rung and one has no effort control at all — and because every model carries its own default, changing model moves the dial even when you do not touch it.

Pick the current model. Then set effort explicitly, because the thing you are really choosing is where the dial starts.