Claude model options: thinking, effort levels and fast mode

1.Extended thinking and adaptive thinking

A model that answers in one pass has to get everything right the first time, with no scratch work. Thinking lifts that limit. With thinking active, Claude works through the problem in its own words before it answers: it restates the question, tries approaches, checks intermediate results and drops paths that fail. This reasoning arrives in thinking content blocks before the text. It is why thinking helps on math, coding, analysis and long agentic tasks. The text you can see in a thinking block is a summary of the reasoning, never the raw chain of thought.

There are two ways to configure it. Extended thinking is the older, manual form: you send type: "enabled" with a fixed budget_tokens. It is deprecated on Opus 4.6 and Sonnet 4.6, and on Claude 4.7 and later models budget_tokens returns a 400 error. Adaptive thinking (type: "adaptive") lets Claude decide when to think and how much, based on the effort setting and how complex the query is. Easy queries get a direct answer. On Claude Opus 5.5, Fable 5.1 and the Fable and Mythos 5 models, adaptive thinking is always on and cannot be disabled.

Thinking is never free. Reasoning tokens are billed as output tokens and count toward max_tokens, so leave room for both the thinking and the reply. The display field only controls whether you see the summary. "omitted", the default on the Claude 5 models and Opus 4.7 and 4.8, returns an empty thinking field and speeds up the first text token when streaming, but you still pay for the full thinking. You can also steer when adaptive thinking triggers. If a large system prompt makes Claude reason through trivial questions, add an instruction telling it to answer directly when thinking would not meaningfully improve the answer.

Sources12

2.Effort levels: how many tokens Claude spends

The effort parameter sets how many tokens Claude spends on a response, so one model can trade thoroughness against token efficiency. You set it with output_config.effort. It affects every output token: the response text, tool calls and their arguments, and thinking when thinking is active. That is why effort works whether or not thinking is on, and why lower effort also means fewer, terser tool calls. On a model where adaptive thinking cannot be turned off, such as Opus 5.5, effort becomes the main control over how much the model reasons and what a request costs.

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=4096,
    messages=[
        {
            "role": "user",
            "content": "Analyze the trade-offs between microservices and monolithic architectures",
        }
    ],
    output_config={"effort": "medium"},
)

Most models default to high. Claude Opus 5.5 defaults to medium, so a request that leaves effort out runs one level lower than it did on Opus 5. Not every model has every level. max and xhigh exist only on newer models, and not every model with max also has xhigh. At the higher levels, set a large max_tokens, because it is a hard limit on total output, thinking plus response text.

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=4096,
    messages=[
        {
            "role": "user",
            "content": "Analyze the trade-offs between microservices and monolithic architectures",
        }
    ],
     ? ={"effort": "medium"},
)

Sources3

3.Fast mode: the same model, faster output

Fast mode is a research preview that gives up to 2.5x more output tokens per second on Claude Opus 5.5, Opus 5 and Opus 4.8, at premium pricing. It is not a different model. It runs the same weights with the same behaviour on a faster inference configuration, so you keep Opus-level capability without moving to a smaller model. You opt in by setting speed: "fast" and sending the fast-mode-2026-02-01 beta header.

response = client.beta.messages.create(
    model="claude-opus-5-5",
    max_tokens=4096,
    speed="fast",
    betas=["fast-mode-2026-02-01"],
    messages=[
        {"role": "user", "content": "Refactor this module to use dependency injection"}
    ],
)

Where the speed shows up is important. Fast mode raises output tokens per second, not time to first token, so it pays off on long generations and is easiest to see when streaming. If your problem is the wait before the first token, fast mode is the wrong tool. Effort and fast mode also act on different things. Effort changes how many tokens Claude produces. Fast mode changes how quickly each one arrives.

Sources4

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.Fast mode swaps in a smaller, faster model, so you trade some intelligence for speed.Why is that wrong?

   Fast mode uses the same model weights and behaviour on a faster inference configuration. Only output speed and price change.

   Covered in Fast mode: the same model, faster output

2. 2.Setting thinking display to "omitted" cuts thinking costs, because the thinking text is not returned.Why is that wrong?

   Omitted display skips streaming the thinking text, which lowers latency, but every thinking token is still generated and billed.

   Covered in Extended thinking and adaptive thinking

3. 3.Effort only matters when thinking is enabled, because it is a thinking budget.Why is that wrong?

   Effort applies to every output token, including response text and tool calls, so it works with or without thinking.

   Covered in Effort levels: how many tokens Claude spends