Domain 3 — Product and Model Selection

CCA Associate Foundations course · Page 4 of 10 · ← Back to all courses · Weight: 12%. Mark this page complete at the bottom to advance your course progress.

3.1 — Match the Entry Point to the User, Not the Capability · Core

The decision rule: name the user before naming the entry point. "Same model underneath" is not a valid reason to recommend the API to a non-technical team — Claude.ai Projects is simpler, cheaper, and immediately usable for a team with no engineering resources. Recommending the API "for everything because it's the most powerful and flexible" is a named exam anti-pattern: it optimizes for raw capability instead of matching the actual user and job.

3.2 — The Three Model Tiers and Their Trade-Off · Core

Default to Sonnet. Move to Haiku once measurement shows the cheaper tier holds accuracy; move to Opus only once evals show Sonnet actually falls short. "Best available model everywhere" is not a strategy — a team that ran Opus on every step of a 5-step pipeline (including simple intent classification) saw 7× cost and 2.3s vs. 800ms target latency, with no change in customer satisfaction.

Routing beats a single model at scale: a Haiku classifier triages requests and escalates only the hard slice to a capable tier — cutting cost with no quality cliff on the easy majority.

⚠ Often-missed — Prompts Don't Transfer 1-to-1 Across Tiers · Gap

A prompt carefully tuned for Opus (heavy scaffolding, many few-shot examples, explicit CoT) is not a finished artifact for Sonnet — a more capable model needs less scaffolding, a less capable one needs more. Every model swap is a release: build a test set with known-good outputs, define a grading function, set the acceptance threshold before running evals, and pin the exact model version in config — never point production at a rolling "latest" alias.

3.3 — Read the Business Constraint First · Core

When two architectures both pass the quality bar and one is cheaper, the cost-constrained scenario always favors the cheaper one — read the named constraint before comparing technical merits.

Extended thinking: available on current Sonnet/Opus tiers via the effort parameter; billed as output tokens and adds latency. Run evals without it first — enable only if accuracy still falls short after prompt improvements. Enabling it "just in case" on every call (including simple routing steps) is a silent cost/latency tax with no benefit.

Batches API: the correct pattern whenever there's no per-request latency requirement — overnight or bulk jobs where cost is the primary concern, regardless of which tier is chosen.

3.4 — Multimodal: What Claude Can and Cannot Do with Images · Core

Claude can read photographs, handwritten forms, charts, screenshots, and scanned documents directly — no separate OCR step needed. It cannot generate images, cannot fetch an image from a URL without a retrieval tool configured, does not provide engineering-grade color measurement (no Pantone-from-photo), and is not a forensic authentication tool. For consequential visual tasks (medical imaging, content moderation), Claude is a first-pass assistant — human review stays in the loop.

3.5 — The Three Parameters the Exam Tests · Core

The context window is one shared budget for input + output tokens together — a 200K-token input leaves zero room for output in a 200K window, regardless of max_tokens. And temperature zero minimizes but does not guarantee identical outputs run to run; design for validated quality, not bit-for-bit reproducibility.

Exam reflexes for Domain 3

• "Non-technical team, repeated weekly workflow" → Claude.ai Projects, not the API.
• "Use the API because it's more powerful/flexible" → wrong; match the user, not raw capability.
• "Cost-constrained scenario, two options both pass quality bar" → pick the cheaper one, always.
• "Opus everywhere, cost/latency exploded, CSAT unchanged" → route by task, don't default to the top tier.
• "Switching model tiers, reusing the old prompt unchanged" → wrong; re-evaluate, prompts don't transfer 1:1.
• "Overnight bulk job, no latency requirement" → Batches API.
• "Output cut off mid-sentence" → raise max_tokens.
• "Same input needs the same output every time" → lower temperature.