Reference

How hard is the CCA-F exam?

Anthropic hasn't published a pass-rate percentage or an official difficulty rating for CCA-F, and this page won't invent one. What follows is structural: what the published scoring formula and domain weights imply, and what the questions are actually shaped like — so you can judge the difficulty for yourself rather than take someone else's unverifiable number.

The pass threshold, in plain terms

CCA-F is scored on a scaled 100–1000 range, with a passing simulation set at ≥720. The scaling formula (as implemented in this site's mock exam, matching the official 60-question / 120-minute format) is:

scaled_score = round(100 + (correct / total) × 900)

Solve that for a 720 result and you need correct / total ≥ ~0.689 — roughly 69% of questions answered correctly. On a 60-question exam, because scores are whole numbers, that works out to needing about 42 of 60 correct (41/60 ≈ 68.3% scores 715 and falls just short; 42/60 = 70% scores 730 and clears it). That's the real bar: not "know most of it," but consistently correct on roughly two questions out of every three, with no domain you can afford to skip.

Difficulty isn't evenly spread across domains

The exam blueprint weights each domain differently, so the stakes of missing an entire topic are not the same everywhere:

D1, Agentic Architecture & Orchestration, carries 27% — nearly double D5's 15% — and is where the exam draws the most questions on when to reach for an agent versus a fixed workflow, how stop_reason should drive control flow, and how multi-agent orchestration and guardrails hold up under failure. A shaky grasp of D1 alone can cost you the several points that separate a 715 from a 730. D2 and D3 (Claude Code configuration, and prompt engineering / structured output) each carry 20% and are close seconds in weight; D4 (Tool Design & MCP) and D5 (Context Management & Reliability) carry less individually but together are still over a quarter of the exam.

What makes the questions hard

The difficulty in CCA-F questions isn't usually obscure trivia — it's scenario-based judgment calls where more than one option sounds defensible. A typical question describes a concrete production situation (an agent looping forever, a tool schema that keeps getting misused, a context window filling up mid-task) and asks for the architecturally correct fix among several plausible-sounding ones. The distractor options are frequently built to look reasonable at first glance: an answer that seems "more sophisticated" but adds unnecessary agentic complexity to a task with fixed, knowable steps; an answer that treats symptom-level patches (like keyword-matching on model output) as if they were the real architectural fix instead of the correct mechanism (like checking stop_reason). Recognizing why the tempting wrong answer is wrong is usually the actual skill being tested — this site calls those out explicitly as exam traps at the end of every lesson.

How this site's structure targets that

The three-stage structure exists specifically to build the kind of judgment the exam tests, not just topic recall:

1. Lessons (29 across all 6 domains) build the underlying mental model for each topic, with an Exam Traps section per lesson calling out the specific misconceptions the real exam is likely to probe.
2. Sprint quizzes (10 questions, 20 minutes, sampled per-domain or mixed) give fast, low-stakes reps at the scenario-question format with an explanation after every answer — good for catching gaps in a single domain before they compound.
3. The full mock exam (all 60 questions, 120 minutes, scored on the same 100–1000 scale with the same ≥720 threshold) is the closest rehearsal for the actual test: no feedback until the end, a per-domain breakdown afterward so you know exactly which domain's weight you're leaving points in, and a full explanation on every question in the review.

Given the weighting, a reasonable prep priority is D1 first and most thoroughly, then D2 and D3 with comparable depth, then D4 and D5 — but every domain needs to clear the bar, since the pass threshold is on the exam total, not a per-domain minimum.