Skip to content

Model Selection Cheat Sheet

As of 2026-10-24. Model names, prices and limits change often. The table below reflects the Claude API reference material cached on 2026-10-06 and current on 2026-10-24. Always confirm on the official models overview and pricing page before you budget or ship. Concepts: 05 · Model selection.

1. Current models (Claude API, first-party prices)

Model API model ID Context window Max output Input $/MTok Output $/MTok Best for
Claude Fable 5.1 claude-fable-5-1 1M 128K $10.00 $50.00 The hardest reasoning and long-horizon agentic work, when cost matters less than capability
Claude Opus 5.5 claude-opus-5-5 1M 128K $4.00 $20.00 Default choice: complex coding, agents, analysis, architecture
Claude Sonnet 5.5 claude-sonnet-5-5 1M 128K $2.00 $10.00 Everyday coding, agents and enterprise workloads with a speed/cost balance
Claude Haiku 5.5 claude-haiku-5-5 1M 128K $0.10 $0.50 High-volume, latency-sensitive tasks: classification, extraction, routing, sub-agents

MTok = one million tokens. Haiku 5.5 prices apply to prompts up to 100K tokens ($0.50 / $2.50 beyond). Previous-generation models (for example claude-opus-5, claude-sonnet-5, claude-opus-4-8) are still served; see the models overview for their status and retirement dates. Use IDs exactly as written: no date suffixes.

Prices on Amazon Bedrock and Google Vertex AI are set by those platforms; Microsoft Foundry bills at Anthropic API rates through its marketplace. Check the provider's pricing page.

2. Price modifiers

Modifier Effect
Message Batches 50% off input and output; asynchronous, up to 24 hours
Prompt cache write About 1.25x base input (5-minute TTL) or 2x (1-hour TTL)
Prompt cache read About 0.1x base input (0.05x on Claude Opus 5.5, 0.025x on Claude Fable 5.1)
Fast mode (research preview) Claude Opus 5.5 at $8 / $40 per MTok on the Claude API only; higher output speed
Effort level Lower effort = fewer thinking and output tokens; the biggest in-model cost lever

3. Decision guide

flowchart TD
  A[New workload] --> B{Hard reasoning, long agentic runs,\nhigh cost of error?}
  B -- yes --> C[Start with Claude Opus 5.5 at high effort.\nTry Fable 5.1 only if evals show a gap]
  B -- no --> D{High volume or\nlatency-critical?}
  D -- yes --> E[Claude Haiku 5.5\nlow or medium effort]
  D -- no --> F[Claude Sonnet 5.5 or Opus 5.5\nat medium effort]
  C --> G[Measure on your eval set]
  E --> G
  F --> G
  G --> H{Quality bar met at\nlowest cost per completed task?}
  H -- no --> I[Raise effort, then change model]
  H -- yes --> J[Ship and monitor]

4. Rules of thumb

  1. Measure, do not guess. Run the same eval set across candidate models and effort levels. Judge by cost per completed task, not per request: a cheap model that needs retries can cost more.
  2. Tune effort before switching models. The most capable model at lower effort often matches a smaller model at higher effort, and keeps one prompt cache.
  3. Route by task, not by team. A support pipeline can use Haiku to classify, Sonnet to draft and Opus to handle escalations.
  4. Use the right model for sub-agents. Orchestrator on Opus, reading-heavy workers on Haiku or Sonnet.
  5. Batch anything that can wait. Nightly summarisation, back-fills and evaluations belong in Message Batches.
  6. Respect data-retention rules. Some models (for example Claude Fable 5.1) require 30-day data retention and are not available under zero data retention unless Anthropic authorises it.
  7. Plan migrations. New models change behaviour and sometimes reject old parameters. Re-run evals, re-baseline token counts with count_tokens, and read the migration guide.

5. Matching model to task (examples)

Task Suggested starting point Why
Classify 2 million support emails overnight Haiku 5.5 + Batches Volume, latency tolerant, simple labels
Live chat assistant for an internal help desk Sonnet 5.5, low or medium effort Responsive, capable enough for multi-turn
Refactor a large codebase with an agent Opus 5.5, high or xhigh effort Long-horizon tool use, cost of error
Draft a regulatory impact analysis for review Opus 5.5, high effort Careful reasoning, human review afterwards
Extract fields from scanned invoices Haiku 5.5 or Sonnet 5.5 + structured outputs Schema-bound extraction; test on real samples
Research agent that reads hundreds of sources Opus 5.5 orchestrator, Haiku 5.5 workers Split reading cost from synthesis quality

6. Claude apps (claude.ai) users

In the Claude apps you pick a model from the model selector, subject to your plan and your organisation's settings. The same trade-off applies: larger models for harder thinking, smaller ones for quick tasks. CCAO-F objective CCAO-F-3.2 and CCAO-F-3.3 cover this; see 05 · Model families.

Further reading