Which Claude Model Should You Use? Haiku, Sonnet, Opus or Fable

Start with Haiku 5.5 for bounded, high-volume work, Sonnet 5.5 for well-scoped everyday tasks, Opus 5.5 for complex coding or knowledge work, and Fable 5.2 for unusually open-ended technical work where its additional cost is justified. These are routing hypotheses, not universal rankings.

A fixed-category support classifier is a sensible Haiku pilot. A task that compares conflicting policy documents or produces a multi-step code change is a stronger Opus candidate. Fable is worth evaluating when the work is unusually open-ended and orchestration or frontier technical capability could materially improve the result. Escalate only when a defined validation or quality gate fails.

The decision also depends on input and output length, reasoning effort, caching, tools, retries, latency, deployment surface, and the cost of an error. The comparison below reflects official information checked on November 11, 2026. Pricing, availability, model identifiers, and benchmark results can change, so verify the linked pages before implementation.

Compare Claude models by job, access, and price

Anthropic’s standard API rates below are per million tokens. Haiku has different rates for prompts above 100,000 tokens, and fast modes are separately priced options rather than part of the standard rates.

Model	Starting use	Standard API input / output	Tradeoff
Haiku 5.5	High-volume classification, routing, summarization, and focused tasks	$0.10 / $0.50 for prompts up to 100,000 tokens; $0.50 / $2.50 for longer prompts	Low standard rates suit bounded work, but ambiguous cases need testing.
Sonnet 5.5	Well-scoped writing, document work, routine fixes, and structured extraction	$2 / $10	A practical middle option when a task needs more nuance than a simple classifier.
Opus 5.5	Complex coding, computer use, document analysis, and knowledge work	$4 / $20	Higher standard rates are appropriate when complexity or error cost warrants them.
Fable 5.2	Ambitious, open-ended technical work or orchestration	$8 / $40	Evaluate its added capability on your own workload before making it a default.

Anthropic lists Opus 5.5 fast mode at $8 input and $40 output per million tokens. Fable 5.2 fast mode is listed at $16 input and $80 output per million tokens. Cache behavior, reasoning effort, generated output, tools, and retries can change the cost of a complete task. Compare the same representative workload, including unsuccessful runs, rather than comparing token rates alone.

See Anthropic’s Haiku page, Sonnet 5.5 announcement, Opus page, and Fable and Mythos announcement for current details. Anthropic also reports that Haiku 5.5 costs less than Haiku 4.5 on stated workloads and that Opus 5.5 costs less to run than Opus 5 on typical workloads. Treat those as vendor comparisons, not guaranteed savings for every workflow.

Anthropic describes Fable 5.2 and Mythos 5.2 as the same underlying model with different safeguards. Fable is generally available, while Mythos access is restricted through cybersecurity and life-sciences verification programs or qualifying enterprise offerings. Availability can also vary by product, plan, platform, and deployment. Do not assume that a context window, maximum output, or plan limit applies identically across Claude.ai, the API, Claude Code, and cloud-provider deployments.

A practical decision matrix for common work

Task	Start with	Escalation signal	Keep deterministic
Stable classification or routing	Haiku 5.5	Ambiguous examples fail review, or valid structured output is unreliable.	Allowed categories, required fields, routing rules, and duplicate checks.
Routine writing or extraction	Sonnet 5.5	Important ambiguity remains unresolved or evidence is incomplete.	Word limits, required sections, exact formats, and thresholds.
Complex coding or document analysis	Opus 5.5	Tests fail, evidence is unsupported, or the task remains unusually open-ended.	Permissions, test execution, acceptance criteria, and release approvals.
Open-ended technical work	Evaluate Fable 5.2	A measured pilot shows a material benefit over Opus for this task.	Tool scope, budget, approval gates, and destination controls.

Anthropic recommends starting with Opus 5.5 for most coding and knowledge work, then moving up to Fable or down to Sonnet or Haiku according to observed need. That is a starting hypothesis, not a reason to send every job to Opus. Exact date arithmetic, known mappings, permissions, monetary thresholds, required-field checks, and duplicate detection are usually better handled by application logic. Use a model where meaning, ambiguity, or semantic comparison needs interpretation.

Turn a model choice into a reliable workflow

Consider a proposed support-classification workflow. A support platform or API emits a record with a stable source ID, URL, text, allowed category list, and prompt version. The application sends the bounded text to Haiku 5.5, parses the response, validates the category and evidence, and then either writes a classification or routes the case to support operations. This is an implementation pattern, not an Anthropic or Zapier template.

A proposed output contract could look like this:

{
  "model_run_id": "run-20261111-000184",
  "source_system": "support-platform",
  "source_record_id": "ticket-8492",
  "classifier_version": "support-classifier-v3",
  "category": "Billing",
  "confidence": 0.91,
  "rationale": "The customer reports a duplicate subscription charge.",
  "evidence": [
    {
      "citation_id": "ticket-8492:0:47",
      "start": 0,
      "end": 47,
      "text": "I was charged twice for the same subscription."
    }
  ],
  "needs_review": false,
  "run_status": "validated"
}

The application should reject invalid JSON, missing identifiers, categories outside the allowlist, and evidence that cannot be matched to the submitted text. Treat confidence as a model signal rather than a calibrated probability unless it has been calibrated against labeled examples. Preserve any existing human decision unless policy explicitly permits replacement.

Keep records at their proper grain. A model-run record represents one execution and stores the model ID, prompt version, effort, token use, tools, latency, and validation result. A classification record represents one source record and classifier version. A citation record represents one evidence span. A pilot summary aggregates defined runs and must not reuse a model-run or classification ID.

For concurrent workers, do not rely on a read-then-create check. Define the intended natural key, such as source system plus source record ID plus classifier version, and enforce it with a database unique constraint or transactional upsert. A retry should reuse the operation key and distinguish a duplicate-key result from a transient failure.

For document comparison, preserve the exact document IDs, versions, retrieval timestamps, URLs, and extracted text. Require source spans for each claimed contradiction and distinguish a true conflict from differences in date, scope, or definitions. Store citation-level evidence separately from the model-run record. Route consequential findings to the relevant policy, legal, or compliance reviewer.

For coding, provide a repository reference, permitted paths, acceptance criteria, test commands, and a maximum runtime or cost budget. Run tests independently, inspect the actual diff for out-of-scope changes, and require human approval before merge or deployment. A model’s report that a test passed is not proof that the test ran.

Model selection is only one part of an agent workflow. Permissions, validation, evidence provenance, destination checks, and a named owner for exceptions matter just as much. See AI agent design and implementation for related service information.

What AutomationBench can and cannot tell you

Zapier’s AutomationBench evaluates end-to-end workflow execution with 47 tools across six business functions in isolated environments. Scoring checks the final environment state against fixed criteria rather than using an LLM judge. Zapier says its task patterns draw on more than 2 billion monthly tasks across 3.7 million companies. That describes the benchmark’s design inputs, not a representative sample of every production workflow.

Anthropic reports AutomationBench results of 51.4% for Fable 5.2 and 42.5% for Opus 5.5. These are vendor-reported, rounded, configuration-specific figures. The live leaderboard can change with model versions, effort settings, safeguards, benchmark versions, and cost assumptions. A benchmark aggregate cannot predict the result of an individual production run.

For an internal pilot, use a fixed, versioned task set and record each run separately: model ID, prompt version, effort, input and output usage, tool calls, latency, validation result, failure severity, and human disposition. Compare cost per accepted result and review effort, not only pass rate or token price. Change one variable at a time where practical.

Using Claude with Zapier: distinguish the product surfaces

Zapier MCP connects compatible AI clients, including Claude, to configured Zapier app actions across more than 9,000 apps. Zapier’s current help documentation says each successful MCP tool call consumes two tasks and failed calls do not consume tasks. Billing and product behavior can change, so check the current MCP usage guidance before budgeting.

AI by Zapier is a separate product surface with its own model tiers and task multipliers. Its current documentation describes 1x, 3x, and 5x tiers. That is not a universal mapping from Claude model name to task charge, and it is not Anthropic API token pricing. See AI by Zapier tier pricing for current terms and model availability.

For a proposed CRM update, pass a stable record ID, restrict the permitted action, verify the target and authorization before an externally visible or destructive write, and inspect the destination record and action history afterward. Use application-level idempotency where available. MCP does not itself guarantee business validation, approval, retries, duplicate prevention, or transactional writes. For help designing a wider automation process, see Zapier automation consulting.

A simple selection rule for your pilot

Define the task, error cost, acceptance criteria, and escalation condition. Choose the least costly plausible model, test it on representative examples, validate both the output and downstream state, and escalate only when a defined gate fails. Re-test when the model, prompt, tools, permissions, or destination behavior changes.

Before relying on an API identifier, check Anthropic’s model lifecycle documentation. Treat Anthropic’s relative cost and capability statements, and Zapier’s live benchmark results, as evidence to evaluate rather than promises about your workload.