basforge

Search titles, text and the “If it doesn’t work” cases.

    RU

    Match reasoning to the task: an effort level instead of “think harder”

    Set reasoning depth with the effort level, not with words in the prompt: low for translation, extraction and chat, high and up for bug hunting, architecture and tasks whose answer can be checked. Reasoning is billed as output and makes the answer slower, so max on everything is wasted money and minutes.

    How it works

    A reasoning model writes a draft before answering: it breaks the task down, tries options, checks itself. The draft is ordinary output tokens, just not shown in full. Models were trained so that on tasks with checkable answers (code, math, logic) the draft noticeably raises accuracy.

    Level When What happens
    low translation, summarizing, field extraction, classification, chat almost no draft: fast, cheap, shorter answers and fewer tool calls
    medium typical code, explaining a config, reading a log the default on Opus 5.5 and Haiku 5.5
    high finding a non-obvious bug, a change across several files the default on Sonnet 5.5
    xhigh long agent work on code between high and max
    max when correctness matters more than cost: architecture, migrations, hard math thinks the longest; worth it only if xhigh visibly falls short

    What changed in the 5.5 models

    • Opus 5.5 can’t turn reasoning off: thinking: {type: "disabled"} returns 400. The minimum is effort: "low".
    • Sonnet 5.5 can turn it off only this way: thinking: {type: "between_tools"}, and only with effort no higher than high.
    • The old method — a fixed budget_tokens budget — is gone on 5.x models; a request with it gets 400.
    • The reasoning text isn’t returned by default (display: "omitted"): the thinking block is there but empty. For a summary of the reasoning, use display: "summarized". You pay the same either way.

    What it costs

    An answer of 500 tokens of text, with 6,000 tokens of reasoning. On Opus 5.5 ($20 per MTok of output):

    without reasoning:   500 × $20 / 1M = $0.010
    with reasoning:    6,500 × $20 / 1M = $0.130   ← 13 times more, and tens of seconds slower

    On “find out why the build fails” that pays for itself with one right answer instead of three wrong ones. On “translate this paragraph”, it doesn’t.

    Why not with words

    “Think carefully”, “this is very important”, “think step by step” change almost nothing on reasoning models: effort sets the depth. Words set something else — what to check before answering: “before you answer, check that every key exists in the Traefik 3.7 documentation in the <docs> block”.

    Where to switch it

    Where How
    API output_config: {effort: "low"}
    claude.ai the button with the model name and level under the input field (Sonnet 5.5 Medium): the menu has the model and an Effort item
    Claude Code the /effort command, the Effort slider in /model; in the VS Code extension, the model button in the input field (Opus 5.5 High), with a slider at the bottom of the model list

    Compare on your own tasks: five typical requests on low and on high. If the low answers are no worse, keep low for that kind of task (Price of a solved task).