Skip to content
AI Product Engineer Wiki

    ↑ ↓ to move between results, Enter to open, Esc to close

    Menu

    Model selection cheat sheet

    Current Claude models side by side, with prices, limits and which kinds of feature each one suits, generated from the wiki's model config.

    1 min read Updated
    On this page

    Current models

    Generated from the same configuration file every code example on this site uses, so it’s always consistent with the examples:

    • Claude Fable 5.1

      claude-fable-5-1

      Context
      1M
      Max output
      128K
      Per 1M tokens
      $10 in / $50 out

      The hardest reasoning and long-running agent work, where quality matters more than cost.

    • Claude Opus 5.5

      claude-opus-5-5

      Context
      1M
      Max output
      128K
      Per 1M tokens
      $4 in / $20 out

      The default for product features: complex tasks, agents and anything quality-sensitive.

    • Claude Sonnet 5.5

      claude-sonnet-5-5

      Context
      1M
      Max output
      128K
      Per 1M tokens
      $2 in / $10 out

      Everyday features at high volume where speed and cost matter.

    • Claude Haiku 5.5

      claude-haiku-5-5

      Context
      1M
      Max output
      128K
      Per 1M tokens
      $0.10 in / $0.50 out

      For prompts up to 100K tokens; $0.50 input and $2.50 output per million beyond that.

      Classification, routing, extraction and other small, latency-sensitive steps.

    First-party Claude API list prices, checked on October 17, 2026. Prices change: see the pricing page (opens in a new tab) and models overview (opens in a new tab) before budgeting. "Good for" is the wiki's recommendation, not a benchmark.

    A decision guide

    Which model to start with. Start capable, then trade down with evidence from your evals.
    Show text descriptionHide text description

    A sequence of yes-or-no questions, asked in order. A "yes" answer stops with the outcome shown; a "no" moves to the next question.

    1. Is this the hardest reasoning or a long-running agent task where quality matters most? If yes: Start with the most capable tier, and measure whether it earns its cost.
    2. Is it a simple, high-volume step such as classification, routing or extraction? If yes: Try the smallest, fastest model, and verify on your eval set.
    3. Is it high volume with moderate difficulty, where speed and cost matter? If yes: Compare the mid tier with the default at low effort on your evals.

    If the answer to every question is no: Use the default model at a sensible effort, prove the feature works, then optimize.

    Then follow the process in choosing a model: lower the effort before switching models, compare settings on your own eval cases, and rerun the comparison when models or prices change.

    Effort in short

    Effort Typical use
    low Classification, extraction, short rewrites, chat
    medium Most product features; a sensible starting point
    high and above Hard reasoning, complex code, long agent tasks

    Set it per request with output_config: { effort }. Thinking tokens count towards max_tokens and are billed as output, so leave headroom in max_tokens on models that think by default.

    Stop reasons

    Every response says how it ended. Handle every value; the error, refusal and empty states article has a typed table for the interface.

    stop_reason Meaning What to do
    end_turn The model finished its answer. Show the answer as complete.
    max_tokens The answer hit the max_tokens limit and is cut off. Label it as cut off; offer a shorter answer or Continue. Never run a tool call from it.
    stop_sequence The model produced one of your custom stop sequences. Treat as complete if you set the stop sequence on purpose.
    tool_use The model wants you to run a tool and send back the result. Validate the input, run the tool (with approval if risky) and send all results back.
    pause_turn A long-running turn paused; send the response back to let it continue. Send the response back unchanged so the model can continue.
    refusal A safety classifier declined the request; stop_details says why. Show a neutral message and a manual path; consider server-side fallback.
    model_context_window_exceeded The conversation filled the context window, so the answer is cut off. Treat like a cut-off answer; reduce context or split the task.
    compaction Only with the compaction beta: the API returned a compacted summary of the conversation. Only with the compaction beta: keep the compaction block in the history.

    Things to check before relying on a number