Claude Certification Blog
Model selection and optimization: the CCDV-F domain that barely selects models
One CCDV-F domain is named after the smallest skill inside it. Plan from the domain title and you spend a study block on 2.7% of the paper while missing the 6.1% sitting beside it.
Model Selection and Optimization is the second-largest domain on CCDV-F at 16.8% — about 9 of 53 items. It contains four published skills, and the one the domain is named after is the smallest of them: Model Selection and Tradeoffs, at 2.7%. The largest is Technical Fundamentals at 6.1%, followed by LLM Fundamentals at 5.2% and Cost and Token Management at 2.8%.
The name and the weights disagree
Domain titles are summaries and always have been. What makes this one worth writing about is that CCDV-F is the only track of the four that publishes weights at skill level, so it is the only exam where the gap between a domain’s name and its contents can be measured rather than suspected. The same effect almost certainly exists on the other three; there is simply no published number to catch it with.
The four skills
| Skill | Weight | What it covers |
|---|---|---|
| Technical Fundamentals | 6.1% | SDKs over REST, and when a held connection beats polling |
| LLM Fundamentals | 5.2% | Tokens, context, sampling, non-determinism, shot count |
| Cost and Token Management | 2.8% | Budgeting, usage tracking, cost modelling, caching |
| Model Selection and Tradeoffs | 2.7% | Capability tiers against quality, latency and cost |
Plan the block, not the title
Four study days on this domain should go roughly two to the two fundamentals skills and one each to cost and to selection — not four days comparing model tiers. The title tells you what the domain is about; only the skill table tells you what it is worth.
Technical fundamentals, the biggest one
At 6.1% this is the largest skill in the domain and the one least likely to appear in anybody’s revision notes. It covers integrating through SDKs that wrap REST APIs, and the choice between holding a connection open for a result and polling for it.
The examinable judgment is what the SDK is doing for you that you would otherwise maintain: request construction, retries, connection handling, response parsing, and keeping up with an interface that moves. That is the same layer question examined across Agent SDK, Claude Code or the API, arriving here from the optimisation side rather than the architecture side.
LLM fundamentals
The 5.2% skill, and the most conceptual material on the paper: tokens, the context window, sampling and non-determinism, the model options including reasoning effort, and prompting with zero, one or several examples.
One consequence carries more exam weight than the rest. Because the same input can produce different output, correctness cannot be asserted by comparing against an expected string — which is why the exams keep arriving at measurement rather than at exact matching. That thread runs through writing evals and evaluating output on every track.
The choosing part, all 2.7% of it
Roughly one or two items. What they ask is not which model to pick but what you gave up to pick it: capability tiers weighed against quality, latency and cost, and the fact that behaviour changes across model releases so a choice made once is not permanently correct.
That last point is why memorising a ranking is a poor investment and writing down a bar is a good one. A configuration chosen against a stated threshold survives a model release; a configuration chosen because it was the recommended one does not. The trade itself is covered in choosing between Claude models, which stays deliberately clear of naming a current lineup.
Cost and tokens
The remaining 2.8%: token budgeting, usage tracking, cost modelling and caching. Small, and unusually easy to secure, because the judgments are concrete rather than contextual.
The one worth internalising is what a cache rewards. Reuse across requests is the whole mechanism, so a prompt whose stable material leads and whose varying material follows is cheaper than the same prompt assembled in the order it was written. Ordering is a design decision with a price attached, which is also why context management and cost control keep turning out to be the same subject.
Key takeaways
- 16.8% domain, 2.7% named skill. Model Selection and Tradeoffs is the smallest of the four skills in the domain that carries its name.
- Technical fundamentals is the largest. At 6.1%, SDKs over REST and the held-connection-versus-polling choice.
- Non-determinism drives the rest of the exam. The same input can differ, so correctness is measured rather than string-matched.
- Write a bar, not a ranking. Behaviour changes across releases, so a threshold survives where a remembered recommendation does not.
- Caching rewards ordering. Stable material first, varying material after, and the cost difference is real.
- Only CCDV-F publishes skill weights. The same title-versus-contents gap probably exists elsewhere and cannot be measured there.
Skill weights are only useful if your practice matches them
A blueprint that resolves to skills rather than domains is a rare advantage, and it is wasted if your papers still allocate by domain title. Timed mocks built to the published 53-item allocation are where you find out whether the block you spent four days on was worth four days. Our claude certification study guide covers how to sequence them, and the domain that is a third of this exam covers where the rest of the time belongs.
See the CCDV-F blueprintQuestions
Frequently asked
The follow-up questions people search next.
How much of CCDV-F is model selection?
The domain called Model Selection and Optimization is 16.8% of the paper, about 9 of 53 items. But the published skill named Model Selection and Tradeoffs is only 2.7% — roughly one or two items. The rest of the domain is technical fundamentals, LLM fundamentals, and cost and token management.
What are technical fundamentals on the Claude developer exam?
At 6.1% it is the largest skill in the domain. It covers integrating with SDKs that wrap REST APIs, and the choice between holding a connection open and polling for a result. Nothing about it is specific to any particular model.
Do I need to memorise which Claude model to use for which job?
Not as a lookup table, and doing so would age badly. What is examined is the trade being made — capability against quality, latency and cost — and the fact that behaviour changes across model releases, which is why a configuration should be chosen against a written bar rather than a remembered ranking.
Is prompt caching examined?
It sits inside Cost and Token Management, a 2.8% skill alongside token budgeting, usage tracking and cost modelling. The examinable judgment is about what a cache rewards — a stable block that leads, reused across requests — rather than about configuring one.
Why does the domain name not match its contents?
Domain titles on every track are summaries, not contents lists. CCDV-F is the only exam that publishes weights at skill level, which is the only reason the mismatch is visible here. On the other three tracks the same effect exists and cannot be measured.
Keep reading
Related posts
Not affiliated with, or endorsed by, Anthropic or Pearson VUE. Details are summarised from publicly published program information and can change — always confirm against the official exam guide before booking.