CCDV-F Claude Certified Developer-Foundations Questions and Answers

Option D treats prompts as production configuration and executable application behavior rather than informal text. A prompt modification can change output format, tool-selection behavior, policy enforcement, downstream compatibility, token usage, or model quality just as materially as an ordinary source-code change. Consequently, prompt changes need traceability, review, testing, and rollback.

A single source of truth prevents different application paths from silently using inconsistent prompt versions. Version control provides an immutable history showing who changed the prompt, what changed, and which application release used it. Pull-request review enables peer assessment and automated evaluation before deployment. The same configuration-as-code principle appears throughout Claude Code’s project configuration model, where shared project configuration is designed to live with version-controlled project artifacts.

A removes governance and creates an uncontrolled external state. B increases hidden runtime behavior and bypasses normal review. C solves fragmentation but leaves the exact governance failure described in the scenario—the ability to ship prompt changes without review.

Therefore, D provides reproducibility and controlled change management. Relevant Study Guide topics: prompt versioning, configuration management, source control, change review, reproducibility, regression testing, and rollback.

===============

The supplied exam source marks B as correct. The first debugging action should be evidence collection rather than immediately modifying prompts or configuration. A trace exposes the complete conditions under which the refusal occurred: user input, system instructions, accumulated conversation state, tool results, parameters, and resulting model output.

This is important because a refusal can originate from several interacting causes. A particular input pattern may trigger it; system instructions may unintentionally conflict with the requested task; prior conversational context may alter Claude's interpretation; or contextual information may make an otherwise acceptable request appear out of scope. Without examining the failing execution path, changing instructions is speculative.

Option A jumps directly to remediation before identifying the cause. C provides useful aggregate analysis but is more appropriate after individual failing executions have been inspected. D is also valuable for comparative debugging, but determining what changed requires the detailed execution data that traces provide.

The correct troubleshooting sequence is therefore: capture or inspect representative failed traces, determine the trigger, compare against successful interactions, formulate a hypothesis, make a targeted change, and re-evaluate.

Relevant Claude Developer topics: Systems Life Cycle, observability, production debugging, tracing, failure analysis, refusals, regression investigation, and evaluation-driven remediation .

===============

Option A is the appropriate trust-boundary treatment for external content. Untrusted text can contain malformed data, adversarial instructions, prompt-injection attempts, or content deliberately constructed to alter the agent's behavior. It should therefore be validated or screened before inclusion and clearly represented as untrusted data rather than authoritative application instructions.

Anthropic's prompt-injection guidance distinguishes direct attacks from indirect prompt injection, where Claude processes third-party content such as webpages, emails, documents, or tool output containing hostile instructions. Anthropic recommends input validation and screening, least-privilege access, safe treatment of untrusted tool content, and screening content before Claude acts on it. Importantly, Anthropic notes that tool-result content is treated as untrusted data rather than as a reliable place for application instructions.

B erases the trust distinction and exposes the application to instruction/data confusion. C introduces a deliberately weaker security boundary and does not sanitize the data. D prevents legitimate use cases unnecessarily; untrusted does not mean unusable—it means the data must be handled defensively.

Therefore, A correctly combines validation with explicit trust separation. Relevant Study Guide topics: prompt injection, untrusted content, input validation, sanitization, data/instruction separation, least privilege, and defense in depth.

Option C applies the appropriate context-management technique. The issue is not simply whether the older conversation can physically fit into the context window; it is that stale details remain prominent enough to interfere with the model’s current task. Effective context engineering optimizes signal quality, not merely maximum token retention.

Anthropic’s context-editing documentation describes compaction as summarizing accumulated history and replacing the full history with a structured summary when context becomes large. This preserves important task state while substantially reducing low-value detail that can distract later reasoning.

A retains precisely the stale material producing the problem and therefore does not address context drift. B destroys all useful continuity between turns, including valid goals, decisions, and intermediate state. D is similarly excessive because limiting Claude to only the latest turn discards information that may still be required.

Compaction provides the correct middle ground: retain durable conclusions, current objectives, unresolved issues, and other high-signal state while compressing obsolete conversational detail. Relevant Study Guide topics: context engineering, compaction, conversation history, long-running agents, stale context, context drift, and high-signal state preservation.

===============

The architecture is a manager/supervisor—or orchestrator/subagent—pattern with isolated subagent context. The defining characteristics are a coordinating top-level agent, specialized subordinate agents, delegated responsibilities, and distinct tool/context boundaries. The supervisor determines which specialist should act, sequences work where necessary, and integrates the specialists' outputs into the final result.

Anthropic's published multi-agent research architecture uses this same structural principle. Its Research system employs an orchestrator-worker pattern in which a lead agent analyzes the task, creates specialized subagents, delegates separate research responsibilities, and consolidates their findings. Anthropic emphasizes that good delegation requires each subagent to receive a specific objective, task boundaries, expected output, and appropriate tools.

The question additionally specifies that each travel specialist has its own conversation history and tool list. That eliminates B because context is explicitly not shared globally. C would involve fixed sequential processing rather than supervisor-directed specialization. D describes retrieval augmentation, not autonomous subordinate agents.

Therefore, A precisely matches both the hierarchy and context-isolation characteristics described.

The supplied examination set also marks A as the intended answer. Relevant topics: Agent Patterns, manager/supervisor architecture, orchestrator-worker systems, subagents, delegation, specialized tools, and context isolation.

===============

Option C follows the governing constraint in the scenario: data residency is an enterprise requirement, whereas minimizing infrastructure management is a preference. A deployment architecture cannot knowingly violate a mandatory compliance boundary merely to accelerate delivery. The question should therefore be interpreted under its stated assumption that the described BYOC environment satisfies the enterprise residency policy.

Anthropic's current Managed Agents documentation distinguishes Anthropic-managed cloud environments from self-hosted sandboxes running on infrastructure controlled by the customer. Self-hosting keeps tool execution, files, and network egress within customer infrastructure, although Anthropic documents that model-facing tool inputs and outputs can still pass through its control plane; therefore, actual regulatory approval must be based on the organization's precise data-flow requirements.

B and D explicitly defer or violate the existing policy. A unnecessarily introduces a planned migration away from the residency-compliant architecture without establishing that such a migration would be permissible.

Accordingly, C is the intended certification answer. Relevant Study Guide topics: deployment topology, data residency, self-hosting, compliance constraints, infrastructure ownership, security boundaries, and architecture tradeoffs.

===============

D applies configuration as code and is the answer explicitly selected in the supplied examination source. Claude Code treats project configuration as part of the repository's working context. Anthropic documents project-level CLAUDE.md as the mechanism for providing project instructions and context, while shared project configuration belongs in .claude/settings.json. Crucially, Anthropic states that .claude/settings.json is intended to be checked into source control and shared with the team .

Keeping these files alongside the code establishes a single versioned relationship between application behavior and the configuration that influences Claude Code. Standard pull requests then provide diffs, peer review, ownership controls, CI validation, audit history, and straightforward rollback. A particular commit can therefore reproduce both source code and its associated Claude project configuration.

A creates competing copies and synchronization risk. B artificially separates behavior configuration from the code version it affects. C creates the same problem specifically for settings.json.

Developer-local values remain distinct: .claude/settings.local.json exists specifically for settings that should not be committed.

Relevant Claude Developer topics: Confia Management, CLAUDE.md, settings.json, configuration as code, source control, pull-request review, reproducibility, auditability, and team-shared Claude Code configuration .

Option D is correct because the problem is not insufficient nominal context capacity; it is degraded signal quality as the working context grows. Anthropic’s context-window guidance explicitly states that more context is not automatically better. As conversations expand, recall and accuracy can degrade through “context rot,” so long-running agents need active context management rather than simply retaining every prior tool result.

A good pattern preserves the active task state while compacting, summarizing, or pruning stale intermediate outputs. Anthropic documents server-side compaction for long-running conversations and context-editing/pruning approaches that clear old tool results when they no longer contribute useful information. This keeps high-value constraints and conclusions visible without carrying every verbose response forward.

Option A changes frameworks without fixing the architectural cause. Option B may reduce call count but also creates larger multi-purpose tools and does not guarantee better context quality. Option C increases capacity while preserving the same low-signal accumulation, so the relevant facts can remain buried.

Therefore, D best implements context engineering for agentic workloads. Relevant Study Guide topics: context engineering, context rot, compaction, tool-result pruning, long-running agents, state preservation, and context-window optimization.

===============

Option D is directly supported by both the supplied examination material and Claude Code's MCP documentation. The examination source selects STDIO for this local-process scenario.

Claude Code's official MCP guidance distinguishes remote HTTP servers from local STDIO servers . Anthropic states that STDIO servers run as local processes and communicate through standard input and standard output. They are particularly suitable for local tools requiring direct machine access or for custom scripts. Claude Code can launch the process itself and communicate with it without creating, exposing, securing, or maintaining a network listener.

HTTP is appropriate when the MCP server exists as a separately reachable network service, especially for cloud-hosted or shared services. That requirement is absent here: the server is being invoked locally from a Claude Code session.

A creates duplicate transport infrastructure without a stated availability requirement. B incorrectly assumes one transport is universally superior regardless of deployment topology. C describes HTTP polling rather than the normal MCP transport relationship and introduces unnecessary request overhead.

Therefore, local execution strongly favors STDIO; remote/shared deployment generally favors HTTP.

Relevant Claude Developer topics: MCP architecture, STDIO transport, HTTP transport, Claude Code integrations, local process communication, and deployment topology .

===============

The supplied Claude Developer item selects A . The symptoms—duplicated logic, tangled control flow, and increasing difficulty adding features—indicate accumulated technical debt in a component that has become structurally difficult to change. Because the problem is already materially reducing development velocity, deliberate refactoring should occur before additional feature pressure compounds it.

The objective is behavioral preservation with structural improvement. Duplicated prompt construction should be consolidated where the behavior is genuinely shared. Prompt preparation, configuration, API invocation, validation, error handling, and post-processing should have clear responsibilities. Control flow should be simplified so each stage is testable independently. Existing tests should first capture important current behavior so refactoring can proceed without silently changing application semantics.

B can be appropriate for minor cleanup, but the question describes broad systemic complexity. Mixing substantial structural refactoring into unrelated feature tickets makes scope, review, and regression analysis harder. C knowingly allows the debt to compound. D centralizes complexity rather than removing it and violates separation of concerns.

Relevant Claude Developer topics: SW Eng Foundations, refactoring, technical debt, DRY, separation of concerns, maintainability, modular design, regression testing, and control-flow simplification .

===============

Option A correctly separates sampling configuration from context management. Temperature historically controlled the randomness of token selection; it did not increase the number of tokens Claude could accept within a request. Anthropic’s current Messages API documentation continues to describe temperature in terms of randomness and, for newer model generations, marks manual temperature control as deprecated. Therefore, changing temperature cannot solve a context-capacity problem.

Long transcripts instead require context-engineering techniques. Appropriate approaches include chunking documents, summarizing earlier material, retrieving only relevant sections, or using context editing/compaction so high-value information remains visible while unnecessary material is removed. Anthropic’s context-editing guidance explicitly supports summarization and replacement of growing conversation history to keep long-running workloads within usable context limits.

B incorrectly conflates generation parameters with context capacity. C may save some tokens but removes persistent application instructions and is therefore architecturally unsound. D modifies an unrelated parameter without addressing the root cause. Relevant Study Guide topics: context windows, token budgets, sampling parameters, summarization, chunking, and context engineering.

===============

Option D applies the correct hybrid pattern. Anthropic distinguishes workflows—where LLMs and tools execute along predefined code paths—from agents, where the model dynamically determines how to proceed. Anthropic recommends choosing the simplest architecture that meets the requirement: workflows provide predictability and consistency for well-understood tasks, while agents are valuable where flexible model-driven decisions are necessary.

Standard invoice processing is explicitly described as a small number of known steps. That makes a deterministic workflow the appropriate default because the path can be tested, monitored, and reproduced. The unexpected-format branch is different: the system must reason among rerouting, requesting clarification, or human escalation. That decision point is where agentic flexibility adds value.

A makes the entire process autonomous even though most steps do not require autonomy, increasing cost, latency, and unpredictability. B introduces multiple specialized agents despite no demonstrated need for that complexity. C collapses deterministic processing and exception handling into one monolithic model call, making validation and debugging harder.

Therefore, D combines workflow predictability with targeted agentic reasoning. Relevant Study Guide topics: workflows versus agents, routing, exception handling, hybrid agentic architecture, escalation, and complexity minimization.

===============

B is directly supported by both the supplied examination source and Claude Code's configuration model. The source marks the hierarchical CLAUDE.md approach as correct. Claude Code supports instructions at multiple scopes, allowing broadly applicable standards to be separated from project-specific context rather than duplicated across every repository.

Anthropic documents several CLAUDE.md scopes. Organization-managed instructions can apply broadly; user-level instructions in ~/.claude/CLAUDE.md apply across a user's projects; project instructions in ./CLAUDE.md or ./.claude/CLAUDE.md provide repository-specific architecture, conventions, commands, and workflows. Claude Code loads applicable files according to the directory hierarchy, allowing broad instructions and more specific local instructions to coexist.

This arrangement improves maintainability because common coding standards are defined once at the appropriate scope, while each repository retains only the context unique to that project. A documentation website does not automatically inject rules into Claude Code context. C creates inconsistent manual configuration. D improperly couples unrelated repositories to one repository's configuration.

Relevant Claude Developer topics: Confia Management, CLAUDE.md hierarchy, organization scope, user scope, project scope, repository configuration, instruction inheritance, and configuration reuse .

===============

Option B is correct. A large prompt containing overlapping, contradictory, or obsolete instructions creates instruction ambiguity and increases the likelihood that Claude will apply an unintended rule, over-weight irrelevant guidance, or produce inconsistent behavior.

Anthropic's current prompting guidance reinforces this principle: Claude performs best when instructions are clear, explicit, and direct. Prompt design should minimize ambiguity and provide the model with precisely the context and behavioral requirements needed for the task. Refactoring should therefore remove obsolete requirements, consolidate duplicated instructions, resolve contradictions, and establish one authoritative formulation for each behavioral rule. The revised prompt should then be regression-tested against representative inputs.

A makes an already overloaded prompt larger and relies on instruction ordering instead of eliminating conflicts. C addresses only one defect while retaining redundant and obsolete material. D does not inherently solve contradictions and may fragment authoritative instructions across multiple locations.

Therefore, B addresses the root cause. Relevant Study Guide topics: system prompt design, prompt maintainability, instruction clarity, conflicting instructions, prompt refactoring, regression evaluation, and context quality.

===============

The supplied Claude Developer source explicitly marks A . The scenario contains the defining elements of an orchestrator/subagent architecture : decomposition of a larger objective, delegation of independent subtasks to specialized workers, and aggregation of their outputs by a coordinating agent.

This architecture is appropriate when subtasks can be performed with focused context or specialized tools. Instead of forcing one agent to carry every intermediate detail, the orchestrator can formulate assignments, launch suitable subagents, receive condensed results, identify missing information, and synthesize the final research product. This also enables context isolation and potentially parallel execution.

Anthropic's published multi-agent architecture uses this orchestrator-worker approach for research workloads: a lead agent decomposes a query, delegates work to specialized subagents, and then integrates their findings. This is particularly useful where exploration is broad and individual subtasks benefit from independent context windows. The broader model-selection guidance also identifies orchestrator strategies as useful where work can be partitioned among worker models.

B describes storage rather than delegation. C is a context-management technique, not a decomposition pattern. D centralizes all responsibilities in a single loop.

Relevant Claude Developer topics: Agent Patterns, orchestration, subagents, task decomposition, specialization, delegation, context isolation, and result consolidation .

===============

Option B is correct because plugin dependencies and releases must be versioned deliberately if the team wants reproducible behavior and a reliable rollback path. Claude Code’s plugin documentation supports explicit semantic versions in plugin manifests and marketplace entries, and the plugin system uses the resolved version as part of update detection and caching. Version constraints can also keep dependent plugins on a tested compatible range until the team intentionally upgrades.

The incident happened because the team could not determine which plugin version had previously worked. Explicit tracking solves that directly: record the version in project configuration, pin known-good releases where stability matters, test upgrades, and change versions through reviewed configuration updates. This creates traceability and makes regression isolation much faster.

Option A removes useful extension functionality instead of managing it. Option C treats third-party change as uncontrollable even though the platform provides version-management mechanisms. Option D automatically moving everything to the latest release increases change frequency and can reproduce the same regression problem.

Therefore, B is the correct configuration-management response. Relevant Study Guide topics: Claude Code plugins, semantic versioning, dependency constraints, reproducible configuration, controlled upgrades, rollback, and configuration management.

Option C correctly explains a fundamental property of generative language models. The supplied examination material identifies C as the correct answer. Claude generates subsequent tokens probabilistically rather than retrieving a single fixed answer for each prompt. Consequently, identical or highly similar requests can produce variations in wording, ordering, explanation depth, and sometimes substantive details.

Anthropic's API documentation explicitly states that sampling parameters control randomness and, importantly, that even configurations historically using a temperature of 0.0 were not fully deterministic . Current newer Claude models increasingly manage sampling behavior internally, so application design should not assume byte-for-byte identical responses across executions.

Production systems therefore handle variation through architecture: schema validation for machine-consumed output, deterministic business-rule checks, retries where appropriate, evaluations for acceptable behavioral ranges, and application-level safeguards. Temperature adjustment may reduce variation on models that support the parameter, but it does not fundamentally convert an LLM into a deterministic function.

A incorrectly labels normal model behavior as a defect. B confuses network latency with generation variability. D overstates model snapshot behavior; pinning a model prevents silent model-version changes but does not eliminate sampling variability.

Relevant topics: sampling, nondeterminism, validation, retries, model versions, and robust Claude API integration .

===============

Option A is correct because a model upgrade should be treated as a controlled application change, not a simple identifier substitution. Anthropic’s evaluation guidance recommends defining measurable success criteria and running task-specific evaluations that mirror real production behavior, including edge cases. Its model-migration guidance likewise recommends testing replacement models before moving production workloads.

Here, the new model improves the highest-volume task but changes output formatting on structured-extraction prompts. That means the migration has both a quality benefit and a compatibility risk. The correct response is to tighten the output contract, re-run evaluations against the parsing/schema boundary, then deploy progressively with a feature flag and a per-workflow rollback path. This limits blast radius and preserves a known-good recovery option.

Option B pushes an unverified behavior change directly into production. Option C makes downstream parsing permissive, which can conceal schema drift instead of enforcing a stable contract. Option D permanently preserves a fragile implementation and discards the measured quality improvement.

Taking the model version stated in the question as the scenario, A is the correct lifecycle strategy. Relevant Study Guide topics: model migration, regression evaluation, structured output, compatibility testing, progressive rollout, rollback, and production change management.

===============

Option A is correct because the safest guardrail architecture is defense in depth, not dependence on one control. Anthropic’s guardrail guidance explicitly recommends combining multiple safeguards: input screening and validation, hardened system instructions, safe handling of untrusted tool content, least-privilege permissions, output screening, monitoring, and red-team testing. The important engineering property is independence: if one layer misses an attack or unsafe request, another layer can still prevent harmful behavior or block delivery.

Option B places control only at the output boundary. Human review can be useful for high-risk cases, but it does not protect tool execution, data access, prompt injection, or other failures that can occur before final output. Option C relies on a single system-prompt control; system instructions are probabilistic and cannot provide complete enforcement against adversarial or malformed inputs. Option D is weaker still because model-level safety alone does not enforce application-specific policies.

Therefore, A best matches Claude Developer security guidance: layer preventive, detective, and enforcement controls so there is no single guardrail whose failure exposes the application. Relevant Study Guide topics: guardrails, prompt injection, input validation, output screening, least privilege, defense in depth, and production monitoring.

===============

Option C is the natural integration model for an application that already uses asynchronous I/O. Claude API calls are network-bound operations, so blocking the application’s event loop while waiting for responses would reduce concurrency and impair throughput. The Anthropic Python SDK explicitly provides AsyncAnthropic, and its documented usage awaits client.messages.create() directly. It also supports an aiohttp backend when improved asynchronous concurrency is desired.

This means Claude requests can participate in the same cooperative asynchronous execution model as other database, HTTP, or service calls. While one request waits for remote I/O, the runtime can continue serving other work rather than dedicating the event loop to an idle blocking operation.

A can be used when integrating unavoidable blocking libraries into async software, but it adds thread-pool management when an official asynchronous client already exists. B needlessly redesigns a concurrent REST service around a synchronous integration. D is specifically harmful because blocking the event loop prevents normal concurrent request processing.

Therefore, C uses the abstraction provided for exactly this architecture. Relevant Study Guide topics: Claude SDK clients, asynchronous APIs, awaitable I/O, concurrency, REST-service integration, and scalable application architecture.

===============

Option A provides proper environment isolation and credential blast-radius control. Development, staging, and production represent different trust boundaries and should not share the same API credential. If a development machine, CI job, or staging service is compromised, a distinct credential prevents the attacker from automatically gaining the same access to production.

Anthropic's official workspace documentation explicitly recommends using separate workspaces for development, staging, and production. Workspaces can have their own API keys, members, resource limits, and usage tracking, and keys can be scoped to a specific workspace. Anthropic's authentication guidance also instructs developers to store API keys in a secrets manager, rotate them periodically, revoke suspected compromised keys, and use workspaces to scope credentials by project or environment.

B confuses rotation with isolation: rotating a credential that remains shared across all environments does not create independent security boundaries. C deliberately increases blast radius. D merely substitutes one globally shared key for another and therefore has the same architectural flaw.

The supplied exam source marks A as correct. Relevant topics: Confia Management, secrets management, API-key scoping, environment isolation, credential rotation, workspaces, least privilege, and production security.

The examination source explicitly identifies D . Non-determinism is a normal LLM property, but that fact does not justify assuming every inconsistency is unavoidable. Systematic diagnosis begins by capturing the complete execution conditions for successful and unsuccessful runs.

A useful trace should expose the actual user input, system prompt, prior messages or context, tools and tool results where applicable, model identifier, request parameters, output, stop reason, and relevant execution events. Current Anthropic agent infrastructure exposes session, span, and agent events specifically to provide observability into session state and execution progress. This makes comparison possible: the engineer can determine whether apparently "same" calls actually differ in accumulated context, tool state, model configuration, parameters, or external data.

A changes a variable before establishing the cause and, on newer Claude releases, temperature may not even be configurable. B generates additional samples but does not ensure the underlying conditions are understood. C prematurely classifies unexplained behavior as acceptable nondeterminism.

The correct sequence is observe, compare, form a hypothesis, reproduce, modify, and evaluate .

Relevant Claude Developer topics: observability, tracing, debugging, model nondeterminism, request-state capture, failure reproduction, and production diagnostics .

===============

C converts validation failure from an uncontrolled terminal condition into a first-class recoverable error path . The supplied exam item identifies C as correct. If structured output does not meet the application's contract, it should never be forwarded as though valid, but immediate user-visible failure is also unnecessary when bounded recovery is possible.

A robust flow can retry generation, ask Claude to repair the malformed structure using the validation error as feedback, switch to an approved fallback path, or ultimately return a controlled failure if the retry budget is exhausted. The application must cap these recovery attempts to avoid unbounded loops.

Anthropic's Structured Outputs documentation explains that unconstrained model generation can produce parsing errors, missing fields, inconsistent types, or schema violations that otherwise require error handling and retries. Current Structured Outputs can eliminate many schema-level failures through constrained decoding, although exceptional conditions such as refusal or output truncation still require explicit handling.

A violates the validation boundary. B removes a protective control. D transfers an engineering reliability responsibility to end users.

Relevant Claude Developer topics: Claude App Design, structured output, validation, retries, repair loops, fallback logic, bounded recovery, error paths, and resilient downstream integration .

===============

Option A is correct because model selection requires measurable non-functional requirements, not an undefined objective such as “best.” Anthropic's model-selection guidance explicitly frames the decision around capabilities, speed, and cost and recommends identifying those requirements before narrowing the candidate models.

“Quality” must also be translated into task-specific success criteria—for example classification accuracy, extraction correctness, acceptable false-positive rate, or human preference. Latency may be constrained by an interactive SLA, and cost may be defined per request, per user, or per completed task. Once those thresholds exist, representative evaluations can compare candidate models using the actual prompts and data.

B uses a generic benchmark that may not correlate with the application's workload. C chooses a model tier without requirements or evidence. D optimizes only for raw capability and ignores latency and cost, potentially producing an unnecessarily expensive architecture.

Thus, requirements definition precedes technology selection. Relevant Study Guide topics: Understanding Reqs, functional versus non-functional requirements, quality thresholds, latency budgets, cost constraints, acceptance criteria, workload-specific evaluation, and model-selection tradeoffs.

===============

The supplied Claude Certified Developer Foundations source marks C . Model selection is a multidimensional engineering decision. There is no universally correct Claude model independent of workload requirements; the application must satisfy the required capability or quality while remaining within acceptable latency and cost envelopes.

Anthropic's official model-selection guidance explicitly identifies capabilities, speed, and cost as core considerations and recommends testing models against workload-specific benchmarks rather than selecting them from a single metric. The guidance further recommends evaluating actual prompts and data, comparing response accuracy, quality, and edge-case behavior, and then weighing the resulting performance and cost tradeoffs.

Options A, B, and D each establish one or two dimensions as primary and effectively defer the remainder. That can lead to a technically unsuitable model—for example, a cheap model that fails the quality threshold or a high-quality model whose latency makes the user experience unacceptable.

The correct method is to define minimum acceptable thresholds across all relevant dimensions and benchmark candidate models against the actual workload.

Relevant Claude Developer topics: Claude App Design, model selection, capability, quality, latency, cost, benchmarking, workload evaluation, tradeoff analysis, and production optimization .