5.2 Claude Model Family Selection

Claude Model Family Selection (Opus, Sonnet, Haiku)

Exam Blueprint Focus: Model Selection and Trade-offs is its own weighted sub-skill inside the 16.8% Model Selection and Optimisation domain. CCDV-F expects you to match a stated business requirement to the right Claude tier using capability, latency, and unit economics — and to know that Anthropic ships one current lineup at a time, so a design that names a retired model is wrong regardless of how well it reasons.

Architectural Philosophy of the Claude Tiered Hierarchy

Anthropic ships the Claude family in named tiers rather than one monolithic model. The tier names — Haiku, Sonnet, Opus, and the research-grade Fable line — are stable across generations; the generation number attached to each name is what moves. Reading a tier name alone tells you the intelligence-per-dollar bracket. Reading the full model ID tells you exactly which weights you are calling.

Every current model supports the same platform primitives: the Messages API, vision (image input), tool calling, prompt caching, the Message Batches API, and citations. They diverge on three axes:

1. Reasoning depth — how much multi-step deduction the model sustains without drifting.
2. Latency and throughput — time-to-first-token (TTFT) and sustained tokens per second.
3. Unit economics — dollars per million input and output tokens.

---

The Current Lineup

The table below is the operative model comparison for a 2026 CCDV-F candidate. Prices are per million tokens (MTok) on the first-party Claude API.

Two rows on that table decide most exam questions:

• Context window. Claude Haiku 4.5 is the only current tier still at 200K tokens. Sonnet 5, Opus 5, and Fable 5.1 carry the full 1M-token window at standard per-token pricing — a 900k-token request is billed at the same rate as a 9k-token request. A scenario that says "we must hold a 400,000-token codebase in one request" therefore rules Haiku 4.5 out on capability, not on cost.
• effort. On the current adaptive-thinking models, output_config.effort is the primary intelligence/cost dial. Haiku 4.5 does not accept it, so on Haiku you tune cost with prompt size and max_tokens instead.

1. Claude Haiku 4.5 — Volume, Latency, and Guardrails

Haiku 4.5 is the throughput tier. Use it where the unit of work is small, well-specified, and repeated thousands of times per hour:

• Classification and routing. Intent detection, ticket triage, language detection, and the front-door router in a cascading architecture.
• Extraction at volume. Pulling three fields out of a receipt; tagging metadata; reformatting.
• Safety and guardrail passes. A cheap second model that scores a response for policy violations before it reaches the user (see the dual-LLM pattern in the security chapter).
• Reranking. Scoring 50 retrieved chunks for relevance before the expensive model reads the top 8.

Trade-offs: the 200K context ceiling, no effort parameter, extended (manual) thinking rather than adaptive, and no interleaved thinking between tool calls. Haiku 4.5 is not the right host for a long autonomous agent loop that must plan across dozens of files.

2. Claude Sonnet 5 — The Production Default

Sonnet 5 is the model most CCDV-F application scenarios should land on. At $2/$10 per MTok it costs 40% of Opus 5 while carrying the same 1M-token window, the same 128K max output, adaptive thinking, and the full low–max effort range. It is the correct answer whenever a scenario says "high quality, cost matters, and we serve real users":

• Customer-facing chat and support agents with a latency budget.
• Retrieval-augmented generation over large document sets.
• Tool-using agents of moderate depth.
• Code generation and review inside a CI budget.

3. Claude Opus 5 — Complex Agentic and Enterprise Work

Opus 5 is the tier for work where a wrong answer is expensive and the task is genuinely hard: long multi-file refactors, deep architectural planning, orchestrator roles that decompose work for cheaper workers, and analysis where the reasoning chain is the deliverable. At $5/$25 per MTok it is 2.5× Sonnet 5, so route to it deliberately rather than by default.

Opus 5 has one exam-worthy behavioural rule: thinking cannot be disabled at xhigh or max effort. A request that combines thinking: {"type": "disabled"} with those effort levels returns a 400 error.

4. Claude Fable 5.1 — Frontier Reasoning

Fable 5.1 sits above Opus 5 for demanding reasoning and long-horizon agentic work; thinking is always on and cannot be turned off. At $10/$50 per MTok it is the most expensive tier, and its cache reads are priced at 0.025× base input rather than the standard 0.1×. Reach for it when evals on Opus 5 at high effort still fall short — not as a general-purpose default.

---

Production Decision Framework: Four Vectors

Vector 1 — Task cognitive complexity

Ask how many dependent inference steps the task needs. Single-step extraction or classification → Haiku 4.5. Multi-step synthesis with a known shape → Sonnet 5. Open-ended planning, long tool chains, or genuinely novel analysis → Opus 5, then Fable 5.1 if evals justify it.

Vector 2 — Latency SLA

Map the SLA to the tier, then to effort. A sub-second P95 on a keystroke-latency feature points at Haiku 4.5. A conversational assistant at 2–4 seconds points at Sonnet 5 — and on Sonnet 5 you can drop effort to medium or low to buy latency back before you consider changing model. Stream the response (SSE) so the user sees TTFT, not total generation time.

Vector 3 — Unit economics

Compute cost per transaction, not cost per token. A 20,000-token cached system prompt read on every call costs $0.002 per call on Sonnet 5 (20,000 × $2 × 0.1 / 1,000,000) and $0.005 on Opus 5. Multiply by request volume before choosing. Remember the two structural discounts that apply on every tier: cache reads at 0.1× base input and the Batch API at 50% off both input and output.

Vector 4 — Accuracy criticality

Where an error creates legal, financial, or clinical exposure, buy the more capable tier and add a verification pass. The cost of one bad contract summary dwarfs the delta between Sonnet 5 and Opus 5 on a few thousand documents.

---

Multi-Model Architectures: Routing and Cascading

Pattern 1 — Intent routing

A cheap classifier reads the request and dispatches it to a tier-appropriate handler. Claude Haiku 4.5 with a forced tool_choice classification tool is the standard router: it returns a category in ~200ms for a fraction of a cent, and the expensive model only sees the requests that need it.

Pattern 2 — Model cascading (fallback escalation)

Run the cheap tier first, then escalate on a measurable signal — a self-reported confidence score below threshold, a failed schema validation, or a refusal. Cascading pays off when the cheap tier resolves a large majority of traffic; if it only resolves half, you pay twice on the rest and the blended cost can exceed calling Sonnet 5 directly. Measure the deflection rate before committing to the pattern.

Pattern 3 — Orchestrator/worker split

Opus 5 plans and decomposes; Haiku 4.5 or Sonnet 5 workers execute the leaf tasks in parallel; Opus 5 synthesises. You pay frontier prices only for the planning and synthesis tokens.

---

Common Traps & Antipatterns

• Naming a retired model. Claude 3 Opus, Claude 3.5 Sonnet, Claude 3.7 Sonnet, and Claude Haiku 3.5 are retired on the first-party Claude API. An answer choice built on them is wrong even if the reasoning is sound.
• Assuming every tier has a 1M window. Haiku 4.5 is 200K.
• Reaching for a bigger model when the real fix is effort or a better prompt. On Sonnet 5 and Opus 5, raising effort is cheaper and faster to test than a tier change.
• Paying long-context surcharges that no longer exist. The full 1M window is standard-priced on Claude 4.6 and later models.
• Defaulting to Opus 5 "for quality." Sonnet 5 at high effort clears most production bars at 40% of the price.