Claude API Model Guide: Choosing Haiku, Sonnet, or Opus

The Claude API offers three model families — Haiku, Sonnet, and Opus — each tuned for a different point on the speed-cost-intelligence curve. Choosing the right claude api model for your use case is a core skill the CCA-F (Claude Certified Architect — Foundations) exam tests directly, and one every production deployment gets wrong at least once.

What Models Are Available Through the Claude API?

Anthropic maintains a current model catalog on the Models Overview page. As of mid-2026, three families cover the full capability range available through the API:

• Claude Haiku 4.5 — the fastest and most cost-efficient model, designed for high-throughput, low-complexity tasks.
• Claude Sonnet 4.6 — a balanced production model that combines solid intelligence with practical throughput.
• Claude Opus (currently 4.6, 4.7, and 4.8) — the highest-intelligence family, built for complex reasoning and long-horizon agentic work.

Every API call requires an exact model ID string. Understanding what each tier can and cannot do is the foundation of sound architecture — and of the Anthropic Claude Certified Architect exam.

How Do Haiku, Sonnet, and Opus Differ in Capability and Speed?

Claude Haiku 4.5 — Speed and Scale

Haiku is the model you reach for when latency and cost dominate the decision. It handles tasks with modest reasoning demands well: sentiment classification, entity extraction, routing logic, and concise question-answering. One architectural constraint worth knowing for the exam: Claude Haiku 4.5 has a 200K token context window — smaller than the 1M token window available on Sonnet and Opus. When a document or conversation thread exceeds roughly 150K tokens, Haiku cannot process it in a single API request, and your design must account for that ceiling. Every other current Claude model offers the full 1M context; Haiku is the one exception.

Claude Sonnet 4.6 — The Production Default

Sonnet 4.6 is the general-purpose workhorse. It handles multi-step coding tasks, document summarization, structured data extraction, and multi-turn conversation with noticeably higher quality than Haiku at speeds that work in interactive applications. For most production pipelines — including those that integrate external tools or connect to resources via the Model Context Protocol — Sonnet 4.6 is the sensible default before you have evidence that Opus is necessary.

Claude Opus 4.8 — Intelligence at the Frontier

The Opus family targets the hardest tasks. Claude Opus 4.8, the current flagship, supports adaptive thinking — an internal reasoning step before the model produces its final answer — and is designed for long-horizon agentic work, complex multi-step reasoning, and situations where correctness matters more than speed or cost. It is the right choice when a wrong answer carries real consequences and the task genuinely requires sustained reasoning across many steps.

How Do Claude API Credits and Costs Vary by Model?

The Claude API charges by token — separately for input (what you send) and output (what the model generates). Anthropic publishes these rates on the pricing page; the figures below reflect published rates as of May 2026 and should be verified against current documentation before you build a cost model:

• Claude Haiku 4.5 — $1.00 per million input tokens, $5.00 per million output tokens.
• Claude Sonnet 4.6 — $3.00 per million input tokens, $15.00 per million output tokens.
• Claude Opus 4.8 — $5.00 per million input tokens, $25.00 per million output tokens.

Two API-level mechanisms cut effective costs significantly across all tiers. Prompt caching stores a stable prefix of your prompt server-side; subsequent requests that hit the cache pay roughly one-tenth the normal input rate for those cached tokens. The Batches API runs large volumes of non-latency-sensitive requests at 50% of standard pricing. Both are worth understanding for the exam: the CCA-F tests cost optimization as a genuine architectural concern, not an afterthought. The exam name sometimes shows up among developers as the "Claude 101 certification," but the official designation is Claude Certified Architect — Foundations.

When Should You Use Which Claude Model? (Decision Framework)

The most useful question for model selection is: what is the cost of a wrong or shallow answer? When mistakes are easy to catch and volume is high, optimize for Haiku. When quality must hold up across diverse inputs, Sonnet covers the gap. When an error has real downstream consequences or the task demands sustained reasoning, Opus earns its premium.

Reach for Claude Haiku 4.5 when:

• You need sub-second latency for classification, routing, or simple extraction.
• You are processing thousands of documents in batch and cost-per-call is the primary constraint.
• The total context fits within 200K tokens.
• Downstream validation or human review will catch shallow errors before they matter.

Reach for Claude Sonnet 4.6 when:

• You need reliable coding assistance, document analysis, or structured output generation.
• Your pipeline includes tool use, external API calls, or Model Context Protocol integrations.
• You want the best balance of speed and intelligence without paying Opus rates for every call.

Reach for Claude Opus 4.8 when:

• The task involves chained reasoning, multi-step planning, or autonomous tool use over many turns.
• You are building or evaluating a long-running agent where correctness compounds across steps.
• The intelligence gain measurably improves outcomes and the cost delta from Sonnet is justified.

What Do Claude Model IDs Look Like in an API Call?

The model parameter in every Claude API call must be an exact string. Anthropic provides stable alias strings for each family. These are the identifiers you will use in production code and encounter on the CCA-F exam:

import anthropic

client = anthropic.Anthropic()

# Haiku 4.5 — high-volume, simple tasks
response = client.messages.create(
    model="claude-haiku-4-5",
    max_tokens=256,
    messages=[{"role": "user", "content": "Classify this review: positive or negative?"}]
)

# Sonnet 4.6 — balanced production workloads
response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Summarize this document and extract action items."}]
)

# Opus 4.8 — complex reasoning and agentic tasks
response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=16000,
    thinking={"type": "adaptive"},
    messages=[{"role": "user", "content": "Plan and implement a solution for this problem..."}]
)

Model IDs follow the format claude-{family}-{version}, always lowercase with hyphens. Passing a marketing shorthand like "Claude Opus" or a misspelled variant returns a 404 error — the API requires the exact alias string. The thinking parameter with {"type": "adaptive"} enables adaptive reasoning on Opus models; the model decides how much internal reasoning each request warrants before producing its answer. This parameter is not valid on Haiku 4.5.

What Does the CCA-F Exam Test About Claude API Models?

Plinth Prep is an independent study resource and is not affiliated with, endorsed by, or sponsored by Anthropic. The following reflects our reading of publicly available guidance — not insider knowledge of exam content.

The Claude Certified Architect — Foundations exam evaluates candidates on their ability to make defensible architectural decisions using Claude's platform. Model selection is a recurring topic because the right choice affects latency, cost, output quality, and context handling simultaneously — all of which interact in non-obvious ways at production scale.

Based on Anthropic's public documentation and the exam's stated architectural focus, candidates should be prepared to:

• Identify the correct model ID for each family tier and explain what an incorrect string produces at the API layer.
• Explain why Claude Haiku 4.5's 200K context window is architecturally significant, and in which scenarios it becomes a hard constraint that rules the model out.
• Articulate the speed-cost-intelligence tradeoffs distinguishing Haiku, Sonnet, and Opus.
• Describe how prompt caching and the Batches API reduce claude api credits across all three tiers.
• Explain when adaptive thinking is appropriate and what it implies for latency and per-request cost.

The skill the exam rewards is not rote memorization of model names. It is the ability to hear a production scenario — "we process 50,000 short product reviews per day and need responses in under 500ms" — and immediately identify the model that correctly balances those constraints. Develop that decision reflex and the model-selection questions become a repeatable strength.