Source: Matthew Purcell's LinkedIn post (2026-11-17) on Anthropic's release of the Claude Certified Developer Professional (CCDV-P) exam and updated guides for three existing exams; Matthew Hartman's post the same day on the CCDV-P exam guide as a second source. The official guides are on Anthropic's Partner Academy. Figures below are as reported by those two authors.
What changed
Anthropic added a fourth professional-level certification, CCDV-P, and rewrote the objectives of three existing exams. The new exam is live now. The updated versions of the other three take effect on 7 December 2026, so candidates preparing against the current guides have until then.
CCDV-P at a glance
70 scenario-based questions, 60 scored; 120 minutes (an early preview listing had said 135); passing score 720; US$175.
8 domains and 25 skills. No prerequisite, though Developer Foundations is recommended and its scope is assumed.
The shift is from building Claude applications to operating them in production at scale.
Domains as described by both sources:
Production integration and reliability (largest, 20%): for example, fallbacks that tell rate limits apart from refusals, and grounding.
Observability and incident response: SLOs, drift detection, postmortems.
Evaluation engineering: CI/CD eval gates, LLM-as-judge calibration, judge drift.
Guardrails and cost.
Fleets of coding agents (new): parallel worktrees, AI code review, unattended runs.
Hartman describes the guide's sample questions as one shape: a production system, some evidence (traces, judge-agreement numbers, a red-team result), and the question of what to do next. The wrong options are things a competent engineer might plausibly do that don't address the root cause.
Updates to the existing exams
All three move to 70 questions (60 scored), up from 60, 53 and 63, and every objective appears rewritten.
Associate Foundations (CCAO-F): seven domains become six (troubleshooting folds into prompting and configuration). New objectives cover Skills, code-executed analysis, memory controls, incognito chats and effort levels. The AI Fluency Framework (the "4Ds") underpins every domain.
Developer Foundations (CCDV-F): Applications and Integration drops from 33% to 20%. Eval, testing and debugging rises from 2.6% to 10%, and Claude Code from 3.1% to 8%. New: evaluation design (LLM-as-judge, golden sets) and grounding answers with citations. Removed: generic software-engineering skills and third-party frameworks such as LangGraph.
Architect Professional (CCAR-P): Architect Foundations becomes assumed knowledge. Stakeholder communication halves from 14% to 7%. New: choosing a deployment surface (direct API vs cloud platforms) and an agent runtime. More depth on LLM-as-judge bias, model-version migration and idempotent tool retries.
Reading the shift
The common direction across all four exams is evaluation and operations. Eval design, judge calibration and drift appear in three of the four, coding-agent fleets get a whole domain, and framework-specific content is removed. For a staff engineer, the CCDV-P blueprint reads less like a product quiz and more like a checklist for running LLM systems in production: classify failure types before choosing a fallback, gate releases on evals, calibrate and monitor judges, define SLOs, and run agents unattended with review and isolation.
Preparing
Read the official guide first. Domain weights decide where study time goes, and CCDV-P assumes Developer Foundations scope.
Practise on scenario questions that give evidence and ask for the next action. Pure recall questions won't match the exam.
If you hold or plan a current-version exam, the updated objectives only take effect on 7 December; existing prep material matches the current versions until then.
Community practice sets exist from both authors. Hartman reports blind-checking every question with separate AI solvers and discarding disagreements. Treat any unofficial set as practice, not as the blueprint.
Caveats
This is a summary of two practitioners' readings of the guides, not of the guides themselves. Domain names and weights should be checked against the official documents.
Pass rates and the question bank are not public.
Related
R087: Stanford CS329Z, Engineering AI Agents, a free course map.
R095: Learn Harness Engineering, an open course with production teardowns.
R022: a six-stage AI engineering reading list.
R066: TRUST, a statistical acceptance gate (background for eval gates).