Choosing a Claude model: tiers, effort and fast mode

1.The current Claude lineup and what each tier is for

The current lineup has four models that cover a range of capability, latency and price. All four handle the same basics: text and image input, text output, vision and tool use. So the real question is rarely whether a model can do a task. It is how well the model does it, how fast, and what each token costs.

The Thinking column names the way a model reasons before it answers. Adaptive thinking leaves the decision to Claude: for each request it judges whether to think and how much, so a simple factual question can get a direct reply with no thinking at all. Extended thinking is the manual mode: you set a thinking token budget on each request with budget_tokens, and Claude thinks against that budget before answering. Pick a tier with adaptive thinking when requests vary from trivial to hard, and expect to set the budget yourself on Haiku 4.5.

Each tier has its own kind of work. Claude Fable 5.1 is the most capable model. It is meant for agent sessions that run for hours and for multistep deep research carried through to a finished document. Claude Opus 5.5 is where most workloads start: multihour autonomous coding agents, large-scale refactoring, vision-heavy workflows and computer use. Claude Sonnet 5.5 balances speed and capability for everyday code generation, data analysis and agentic tool use. Claude Haiku 4.5 has the lowest latency and price. It fits real-time applications, high-volume processing and sub-agent tasks, and it is the only tier that still uses extended thinking instead of adaptive thinking.

The price differences add up at volume. Fable's output tokens cost ten times what Haiku's do, so a high-throughput job that Haiku handles well saves a lot. Haiku also has smaller limits: a 200K context window and 64K maximum output, where the other three have 1M and 128K.

Sources1234

2.Two ways to pick a starting model

There are two sensible ways to pick a first model, and the task decides which one to use. Start small and step up works well for prototyping, tight latency budgets, cost-sensitive deployments and high-volume tasks that are simple. You build on Claude Haiku 4.5 and only move to a larger model when you find a specific gap in capability. Iteration stays fast and development stays cheap, and for many common applications the small model turns out to be good enough.

Start capable and step down is for complex reasoning, scientific or mathematical work, tasks that need nuanced understanding, advanced coding, and any case where accuracy matters more than cost. Build on Claude Opus 5.5, tune your prompts for it, and check the results against your requirements. Over time, lower the effort or move to a cheaper model. If Opus 5.5 still falls short at xhigh or max effort on demanding reasoning or long-horizon agentic work, the next step up is Claude Fable 5.1.

Evals settle the choice either way. Write benchmark tests from your real prompts and data. Then compare models on accuracy, response quality and edge cases. When you weigh cost, compare cost per completed task, not cost per token. In Anthropic's measurements, each newer model solved at least as many tasks as the one before it, usually for less per solved task.

Sources25

3.Effort: tuning one model instead of switching

Effort sets how many tokens a model spends on a response. It trades thoroughness against token efficiency, and you set it per request as output_config.effort. It has five levels: low, medium, high, xhigh and max. Effort applies to every output token, which includes the text, tool calls and their arguments, and thinking when thinking is running. Lowering it saves time and money but gives up some capability. Raising it to max gives the most capability.

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"},
)

The default level, and the best place to start, depends on the model. Most models default to high, and on Claude Fable 5.1 and Claude Opus 5 that is the right starting point. Claude Opus 5.5 defaults to medium, so start there. On Claude Opus 4.8 and 4.7, xhigh is the best setting for most coding and agentic work, not the default. Claude Haiku 4.5 does not support effort.

Your results tell you which way to move. If cost is too high and quality is fine, lower the effort on the model you already use. If quality falls short after you lowered effort, put it back. If quality falls short and you never lowered effort, try the next tier up at low effort before you commit to a bigger model at its default.

Sources65

4.Fast mode: the same model, faster output

Fast mode runs the same model on a faster inference configuration. It changes the speed of output, not the model. Weights, behaviour and capability stay the same, and output tokens per second rise by up to 2.5x. The gain is in output speed, not time to first token, so it shows most when you stream a long response. You opt in with speed: "fast" and 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"}
    ],
)

Only Claude Opus 5.5, Claude Opus 5 and Claude Opus 4.8 support fast mode. Claude Opus 4.7 rejects it. Fast mode is billed at a premium: on Opus 5.5 it costs $8 per million input tokens and $40 per million output tokens, double the standard $4 and $20. That makes it the opposite of effort. Effort changes how much work the model does and can lower cost. Fast mode keeps the work the same and buys quicker delivery. Use fast mode when a person is waiting on a long generated output and the speed is worth paying for. Use effort when the problem is the bill.

Sources27

Exam traps

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

1. 1.When costs are too high, the first step is to move to a cheaper model.Why is that wrong?

   If quality is acceptable, first lower the effort on the model you already use. Switch models only if your evals show it is needed.

   Covered in Effort: tuning one model instead of switching

2. 2.For high-stakes autonomous coding on Claude Opus 4.8, the default effort or max is the recommended setting.Why is that wrong?

   On Opus 4.8 and 4.7, xhigh (between high and max) is the recommended level for most coding and agentic work.

   Covered in Effort: tuning one model instead of switching

3. 3.Fast mode swaps in a lighter model, so you trade some quality for speed.Why is that wrong?

   Fast mode runs the same model on a faster inference configuration. Capability does not change; only the output rate and the price do.

   Covered in Fast mode: the same model, faster output