Most people find the model picker in their first week, set it to the strongest name on the list, and never touch it again. There is a second setting beside it, and it is usually the one that would have helped.

The idea

Two dials control what you get. The first is which model: a vendor sells a family of them, differing by capability, speed, price, how much text they hold at once, and how recent their knowledge is. The second is effort: how much work the model spends on your request, set separately from the model and offered on most of the line but not the cheapest tier. Depth and speed are both for sale. They are not the same purchase.

code
  DIAL 1: THE MODEL                    DIAL 2: THE EFFORT
  capability / price / window          how hard it works

  Haiku 4.5   $1 / $5    200K          low     fast, shallow
     |                                  |
  Sonnet 5    $2 / $10   1M            medium
     |                                  |
  Opus 5      $5 / $25   1M            high    <- effort's default
     |                                  |
  Fable 5     $10 / $50  1M            xhigh
                                        |
  $ in / $ out per million tokens      max     slow, deep
  context window

  The columns are independent: rows opposite each other are unrelated.
  Not every model has both dials. Effort is not offered on Haiku 4.5.
  Every current model that offers it offers all five levels; older
  models are where the top levels go missing.
  ---------------------------------------------------------------
  Not a dial, costs nothing, gets tried last: rewriting the request.

How it works

Anthropic's current line runs Claude Haiku 4.5, Claude Sonnet 5, Claude Opus 5 and Claude Fable 5, in that order of capability and price, and older versions stay on sale beside them. Checked 13 August 2026 against Anthropic's published API pricing:

  • Haiku 4.5: $1 per million in, $5 out. Its context window is 200,000 tokens where the rest hold a million.
  • Sonnet 5: $2 in, $10 out.
  • Opus 5: $5 in, $25 out.
  • Fable 5: $10 in, $50 out.

Note the top of that list. The strongest model costs ten times the cheapest, and the picker offers no hint that it does.

GitHub Copilot's model picker mixes families: its model comparison page lists Anthropic's Claude models next to OpenAI's GPT tiers, and the same shape holds there (a cheap fast tier, a mid tier, a top tier). Anthropic publishes each model's on its models overview page.

Effort is set per request, and it applies to whichever model you already picked. Anthropic's effort setting runs low, medium, high, xhigh, max, and defaults to high. A lower setting thinks less on an easy input, makes fewer tool calls, and goes straight to the work instead of explaining first. Thinking is billed as output tokens even when you never see it, so depth is bought at the most expensive rate there is.

Where the dial actually is. In Claude Code, type /effort to change it, or set it alongside the model in /model. Thinking can be switched off in /config on most models, though not on Fable 5, and not on Opus 5 once effort is above high. If you use something else, look for it near the model picker before assuming your tool has no such setting.

What to do

  1. Set a mid model as your default and reach up when a job earns it. Anthropic's own Claude Code cost guidance names a usual cause of a surprising bill: the strongest model left on as the default for everything.
  2. When an answer disappoints, try the effort dial before the model dial. Effort buys thinking at the output rate, so it is not free, but a step up on the model you already have is usually a smaller bill than a step up to the next model. The default is already high, so the move is to xhigh or max, and only for the request that needs it.
  3. Match the properties that are not capability. A job bigger than 200,000 tokens rules Haiku out. A question about last spring's release needs a recent knowledge cutoff, and the spread is wide: Anthropic currently publishes May 2026 for Opus 5 and February 2025 for Haiku 4.5, fifteen months apart on the same price list.

Where it breaks

More is not reliably better. Anthropic's effort documentation says plainly that the highest setting can add significant cost for small gains on structured or less demanding tasks, and can tip into overthinking. Turning both dials up is a way to spend more and wait longer for the same answer.

Everything specific on this page expires. Names, tiers and prices moved twice while this section was being written, so build the habit on the shape: cheap and fast at one end, deep and slow at the other, and a second dial travelling the same axis for less money.

Some products decide for you. An automatic mode picks per request and removes the dial, and you are not told which model answered.