Claude Model Selection: CCDV-F Domain 5 Developer Guide

The CCDV-F developer exam tests claude model selection across 16.8% of its 53 items, making Domain 5 (Model Selection and Optimisation) one of the highest-leverage areas in the eight-domain blueprint. Understand cost-quality-latency reasoning well and you bank roughly 9 questions toward the 720 scaled score required to pass. Miss it and you hand back a substantial portion of the exam.

This guide maps what Domain 5 actually tests, how exam scenarios frame the trade-offs, and which complementary skills in other domains intersect with model selection decisions on the CCDV-F.

What does Domain 5 of the CCDV-F exam actually test?

Domain 5 asks a deceptively practical question: given a production workload with specific constraints, which Claude model is the right choice and how should it be configured? The exam does not test recall of release dates or benchmark scores. It tests practical judgment about cost, latency, context window size, output quality, and the techniques that adjust a model's behaviour within those dimensions.

Per the official CCDV-F exam guide (2026-07-08), items across the exam are written as realistic scenarios rather than trivia. You will not see "which model has the largest context window?" as a factual question. You will see a scenario describing 80,000 invoices processed overnight, a constrained cloud budget, and a 94% extraction accuracy requirement, and you need to reason to the correct answer from those constraints alone.

With 16.8% weight on a 53-item paper, Domain 5 contributes roughly 9 questions to your total score. Unlike the CCAR-F architect exam, which draws 4 scenarios at random from a bank of 6, the CCDV-F has no scenario bank: every item is written directly against the domain's task statements, so rote pattern-matching on scenario structures will not help.

What are the Claude model tiers and when does the exam expect you to choose each?

Anthropic positions its Claude models across three main intelligence tiers: Haiku, Sonnet, and Opus. Each sits at a different point on three production-relevant axes.

The positioning above reflects Anthropic's current documentation; for current pricing and context window sizes, consult Anthropic's model overview directly, since figures update across model generations.

CCDV-F scenarios rarely present a single clean constraint. A typical item combines at least two: "latency under 300 ms AND monthly cost under $200" or "95% extraction accuracy AND context windows up to 120,000 tokens." Your job is to identify the binding constraint and eliminate options that violate it before choosing between those that remain.

How does the cost-quality trade-off appear in exam scenarios?

The cost-quality trade-off is the most common framing in Domain 5 items. The exam tests whether you understand that model selection is not a one-way optimisation toward quality: it is a balancing act where overspending on capability is as wrong as underspending on it.

Here is a scenario pattern worth practising:

Scenario: A SaaS company routes 150,000 support tickets per month to department queues. Routing accuracy must stay above 90%. The pipeline runs overnight, so latency is irrelevant. The monthly API budget is $300.

The right reasoning path:

1. Latency is irrelevant, so that axis does not constrain the choice.
2. 90% accuracy for a classification task with well-defined categories is achievable with a well-prompted Haiku model.
3. 150,000 overnight requests can use the Message Batches API for an additional cost reduction.
4. Opus or Sonnet would likely reach higher accuracy, but the target is 90%, not 99%. Over-provisioning wastes budget.
5. Correct answer: Haiku via Message Batches.

If the scenario instead described routing that "requires understanding nuanced customer sentiment and ambiguous intent across multiple languages," the calculus shifts. Task complexity has increased beyond Haiku's reliable ceiling.

Good Prompt Engineering & Structured Output skills amplify model selection reasoning directly. A system prompt with clear category definitions and a handful of well-chosen examples can raise a Haiku model's classification accuracy by several points, often enough to meet a quality bar that would otherwise require Sonnet. The exam expects you to know this interaction and apply it under time pressure.

Does context window size change the model selection decision?

Yes, and this angle appears regularly. Context window capacity is a hard constraint: if a task requires ingesting a 200,000-token document in a single call and a model's window does not accommodate it, that model is ineligible regardless of cost or quality.

The subtler point the exam tests is the cost implication of large-context calls. Every token sent as context is priced as an input token. A workload that consistently sends 100,000 tokens of context per call sees dramatically different monthly costs across model tiers even when output length is identical.

This is where retrieval-augmented patterns become a model selection lever. If you can retrieve the 3,000 most relevant tokens from a 100,000-token corpus instead of sending the full corpus, you shift from large-context pricing to small-context pricing. Context Management & Reliability covers the retrieval-versus-full-context design decision in depth, and Domain 5 of the CCDV-F expects you to recognise when retrieval makes a cheaper model viable.

A typical exam item might describe a legal document review system where each document is 80,000 tokens. The surface answer is "pick Opus with a large context window." The better answer is often: add a chunking and retrieval layer so each model call sees only the relevant contract clauses, then use Sonnet, reducing cost per document without sacrificing the accuracy the scenario demands.

How does model selection work differently in agentic systems?

Single-model selection is straightforward. Selection in a multi-agent pipeline is not, because agents within the same pipeline carry fundamentally different responsibilities, and optimal model choice varies by role.

A coordinator agent handling orchestration logic, deciding which tools to call, and managing error recovery needs capable reasoning. A sub-agent extracting structured fields from a pre-parsed document section does not. Assigning Opus to both wastes budget; assigning Haiku to both degrades reliability at the coordination layer.

The CCDV-F exam tests this explicitly. A typical item describes a pipeline (coordinator plus several sub-agents) and asks which agent's model tier should be upgraded to fix a reliability problem. The correct answer is the agent performing complex judgment, not the one doing deterministic, well-defined transformation.

Agentic Architecture & Orchestration maps the coordinator-subagent responsibility split in detail. You cannot pick the right model for a role if you have not first understood what that role demands. Domain 1 (Agents and Workflows, 14.7%) and Domain 5 overlap in exactly this way, and exam items regularly test both domains simultaneously.

For agentic scenarios, the exam also probes prompt caching strategy. A coordinator that maintains a stable system prompt across thousands of sub-calls can cache that prefix and reduce the effective per-call cost, making a more capable model financially viable at scale.

What optimisation techniques pair with model selection on the CCDV-F?

The exam treats model selection and model optimisation as one domain because the initial model choice is a starting point, not a final answer. Four techniques appear regularly in Domain 5 scenarios.

Prompt caching. For workloads with a stable, long system prompt (agent instructions, a reference document, a fixed few-shot block), Anthropic's prompt caching feature stores the prefix after the first call. Subsequent calls pay a reduced input-token price for the cached portion. This can make Sonnet or Opus economically competitive with Haiku for high-volume workloads sharing a long stable prefix.

Message Batches API. For non-latency-sensitive workloads, batching requests reduces per-request cost at the price of delayed results. Domain 5 expects you to know when that trade-off is acceptable (overnight classification, daily summarisation runs) and when it is not (real-time user-facing responses).

Few-shot calibration. Carefully selected examples in the prompt can bring a cheaper model's accuracy up to the level of a more expensive model on well-scoped tasks. The exam tests whether you know this is possible and, critically, when it is likely to succeed (consistent classification) versus when it will fall short (open-ended generation requiring deep reasoning).

Model routing. Some production systems route requests to different model tiers based on detected complexity. Simple queries go to Haiku; complex queries escalate to Sonnet or Opus. The following is a conceptual illustration of the routing pattern Domain 5 expects you to recognise:

# Conceptual routing pattern -- consult Anthropic docs for current model identifiers
def select_model(complexity_score: float) -> str:
    if complexity_score < 0.4:
        return "claude-haiku"    # fast, low-cost tier
    elif complexity_score < 0.75:
        return "claude-sonnet"   # balanced tier
    else:
        return "claude-opus"     # highest-capability tier

The exam does not test specific model version strings. It tests whether you can identify which tier a given task should land on given its complexity and constraint profile.

Which other CCDV-F domains overlap with model selection?

Domain 5 does not sit in isolation. Three other domains interact with it on the exam, and understanding the overlaps helps you prioritise study time.

A weak Domain 5 score often reflects gaps in prompt engineering or architecture knowledge as much as model-specific knowledge. If your practice exams show accuracy below 70% in Domain 5, check whether Domain 6 and Domain 1 gaps are the underlying cause before focusing narrowly on model tier facts.

How should you approach Domain 5 study before the CCDV-F exam?

Domain 5 rewards scenario practice over concept memorisation. The CCDV-F is a 53-item, 120-minute exam with a passing score of 720 on a 100-to-1000 scale. With 16.8% of the score riding on model selection reasoning, a structured study plan matters.

1. Read Anthropic's model overview to understand current capability positioning across tiers.
2. Practise articulating trade-off reasoning before checking answers: given constraints X, Y, and Z, I choose tier T because it satisfies A while B and C eliminate the alternatives.
3. Study prompt caching and Message Batches API mechanics from Anthropic's API documentation.
4. Work through agentic architecture scenarios to practise per-role model assignment in multi-agent pipelines.
5. Use the platform's adaptive practice engine, which tracks your Domain 5 accuracy against the 0.90 mastery threshold and surfaces items weighted toward your weakest sub-skill clusters.

Domain 5 carries 16.8% of the score. Bringing your Domain 5 accuracy from 60% to 90% is worth more scaled points than the same improvement in Domain 3 (Claude Code, 3.1%) or Domain 4 (Eval, Testing, and Debugging, 2.6%). Study time allocation should reflect the domain weights directly.

AI Skill Certs is an independent prep platform and is not affiliated with or endorsed by Anthropic. All exam facts cited above are sourced from Anthropic's official CCDV-F exam guide (2026-07-08) and the programme's published documentation.