Best Claude Model for Coding: CCDV-F Domain 5 Guide

For most production coding work, the best claude model for coding is Claude Sonnet 5 (claude-sonnet-5). It handles multi-file refactors, API integration, and code review at a speed and cost that scale to real workloads. Reserve Claude Opus 5 (claude-opus-5) for the hardest problems: novel algorithm design, large-scale codebase analysis, or tasks where the cost of an error justifies the premium. Claude Haiku 4.5 (claude-haiku-4-5-20251001) belongs in high-volume, low-complexity pipelines: inline completions, test generation from templates, and classification tasks where sub-second latency is a hard requirement.

That framework is also an exam topic. If you are sitting the Claude Certified Developer, Foundations exam (CCDV-F, $125 per attempt), model routing is examined directly under Domain 5: Model Selection and Optimisation, which accounts for 16.8% of the 53-question paper. As of 3 June 2026, the Claude Partner Network had attracted more than 40,000 partner applicant firms and certified over 10,000 individuals, reflecting a pool of professionals for whom routing decisions are a professional competency, not academic trivia.

How do the CCDV-F domain weights shape study priority?

The CCDV-F is a 53-item, 120-minute exam scored on a 100-to-1000 scale with a passing mark of 720. Its eight domains carry precise fractional weights, and understanding the distribution is the first step of a rational study plan.

Domain 2 (Applications and Integration) is the dominant slice at 33.1%, but Domain 5 is the second largest at 16.8%. Together they account for just under half the paper. A candidate who answers both confidently is well positioned to clear 720. The two smallest domains, Claude Code (3.1%) and Eval, Testing, and Debugging (2.6%), still deserve proportional attention: a few missed questions from a small domain can swing a borderline score.

A proportional study allocation follows directly from the weights. Spend roughly a third of your preparation time on Domain 2, a sixth on Domain 5, and distribute the remainder across the other six domains in proportion. In practice, our adaptive engine handles reallocation automatically once it detects mastery: once you clear the 0.90 mastery threshold on a domain, the engine redirects study time to weaker areas.

Unlike the CCAR-F architect exam, the CCDV-F has no scenario bank. Items are written directly against the skills in each domain, which means the exam covers the full breadth of each domain's task statements rather than drawing from a rotating pool.

What coding tasks suit each Claude model tier?

Domain 5 questions are scenario-driven. Each item describes a workload, a budget constraint, or a latency requirement, then asks which model and configuration fit best. The following table maps common coding tasks to the appropriate tier.

The pattern the exam rewards is consistent: use the least capable model that satisfies your quality bar. Defaulting to Opus for every coding task is treated as an incorrect answer when the scenario specifies a binding cost or latency constraint. The exam does not reward over-engineering, but it does not reward under-engineering either. If a scenario involves a novel, high-stakes algorithm and you route it to Haiku, that is equally wrong.

It is also worth recognising that the exam does not ask you to memorise a fixed lookup table. Real exam items describe contexts: latency budgets, throughput requirements, the stakes of an error, the size of the codebase. You must reason from context to conclusion. The table above is a starting heuristic; the exam tests whether you can apply the underlying reasoning.

How should you route coding requests between models in production?

The practical implementation of model selection is routing by task complexity before invoking the model. A simple approach classifies tasks programmatically and maps them to tiers:

import anthropic

client = anthropic.Anthropic()

MODEL_TIERS = {
    "low": "claude-haiku-4-5-20251001",
    "medium": "claude-sonnet-5",
    "high": "claude-opus-5",
}

def route_coding_task(task: str, complexity: str) -> str:
    model = MODEL_TIERS.get(complexity, "claude-sonnet-5")
    response = client.messages.create(
        model=model,
        max_tokens=2048,
        messages=[{"role": "user", "content": task}]
    )
    return response.content[0].text

For the CCDV-F, the exam tests model-driven vs. pre-configured decision-making: when should routing logic be hardcoded versus delegated to Claude at runtime? Static routing suits well-defined task categories. Dynamic routing suits exploratory agents where task complexity is unknown until execution begins. Knowing which pattern to recommend in a given scenario is a Domain 5 skill.

Domain 5 also tests prompt caching as an optimisation lever. When coding tasks repeatedly use the same large context, such as a full codebase read into a system prompt, caching reduces per-request cost without changing response quality. Per Anthropic's API documentation, prompt caching is available across all model tiers and requires only a structural change to the request.

A second routing decision the exam tests is synchronous versus batch processing. When a coding pipeline analyses thousands of files nightly, batch processing may outperform synchronous calls on both cost and throughput. Identifying when batch is preferable to synchronous is a skill that spans Domain 2 (Applications and Integration, 33.1%) and Domain 5.

How does context length interact with model selection?

Larger context windows allow you to pass entire codebases to a single model call. That capability is not free of cost or trade-off, and the CCDV-F tests both sides.

Context Management and Reliability is Domain 5's natural complement. Candidates who understand how token volume affects per-request cost, and how to compress or summarise context to remain within budget, perform more reliably on Domain 5 questions. A typical exam scenario: given a large codebase and a cost constraint, what is the optimal approach to answering multiple independent questions about it?

The correct answer is typically to use prompt caching, split requests across smaller context windows, or inject summaries rather than re-sending the full codebase on every call. This is the proportionate-fix principle: apply the least expensive technique that satisfies the requirement.

Token budgeting is a related skill. If you know you will make a large number of requests per day to analyse code, you can model your monthly cost using Anthropic's published pricing before committing to a model tier. Domain 5 expects candidates to perform this kind of estimation: given a token volume, a model tier, and a cost target, which configuration meets the budget? This rewards candidates who have practised with the API rather than only read about it.

Attention dilution is the associated risk. As context length grows, model attention across earlier tokens weakens. For coding tasks this matters when critical function signatures or constraints appear early in a long file list. Structuring prompts to place the most important context close to the instruction, or splitting into per-file passes, is the kind of solution that Domain 5 and Domain 6 (Prompt and Context Engineering, 11.0%) jointly reward.

What role does Claude Code play in coding certification?

Claude Code is Domain 3 in the CCDV-F, weighted at 3.1%. That is the second-smallest domain, but it is not entirely separable from model selection. Claude Code uses the model available in the current session by default but supports model overrides and configuration through CLAUDE.md and project-level settings.

For the exam, what matters is less about memorising CLI flags and more about understanding when Claude Code's built-in tools (file reads, shell execution, web search) represent the right interface versus when a custom API integration is preferable. That distinction is also part of Domain 2's integration scope at 33.1%, which is why the two domains are worth studying in parallel.

Prompt Engineering and Structured Output principles apply equally whether you are calling the API directly or working through Claude Code. A well-designed system prompt improves output reliability across both interfaces, and Domain 6 (11.0%) tests that understanding regardless of which delivery mechanism a scenario uses.

How do security constraints affect model selection for coding?

Domain 7 (Security and Safety, 8.1%) and Domain 5 (16.8%) overlap on a specific question: should high-risk coding tasks, such as generating cryptographic primitives or modifying access-control logic, be routed to a more capable model or constrained to a more conservative configuration?

The exam's answer is clear: capability and safety are separate axes. The correct approach is to apply prompt-level constraints and principled tool design regardless of model tier. A well-constrained Sonnet 5 call is safer than an unconstrained Opus 5 call for the same sensitive task. Domain 7 tests whether candidates know to apply safety guardrails at the prompt and tool boundary, not merely by downgrading the model.

Blast-radius reduction is the recurring principle. Tools that expose file writes, shell execution, or network calls should be scoped and validated regardless of which model invokes them. The CCDV-F does not treat Haiku as inherently safer than Opus; it treats well-designed tool boundaries as the primary defence.

Which domain interactions matter most for Domain 5 prep?

Model selection is not an isolated topic on the CCDV-F. The table below shows where Domain 5's reasoning intersects with the other domains you will encounter across the 53-question paper.

Because Domain 2 is the largest slice and model selection is central to application integration, a candidate who can answer "which model, why, and under what constraints" will find that reasoning transferable across more than half the paper.

What prep approach should you take for Domain 5?

We recommend two phases. In the first, read Anthropic's model documentation to understand the current model family and their stated performance characteristics, then map each to the coding-task categories above. In the second phase, work through practice scenarios that pair a technical task description with a cost or latency constraint, and reason through the routing decision before checking the answer.

Our adaptive engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold. Once you clear that threshold on Domain 5, the engine redirects study time to weaker domains. Domain 5 is tractable: the logic is explicit and the trade-offs are well-defined. A prepared candidate can reliably capture most of its 16.8% allocation.

Developer-track prep for the CCDV-F, including adaptive study, Archie tutoring, and practice exams, is live on the platform today. If you are also exploring the architect track, our Claude Certification Concepts library maps 174 atomic concepts to the five CCAR-F exam domains.

AI Skill Certs is an independent prep platform. We are not affiliated with, endorsed by, or approved by Anthropic.