Anthropic is preparing a fifth Claude certification: Claude Certified Developer - Professional. It sits above Developer - Foundations (CCDV-F) on the developer track, and its catalog entry in Anthropic Partner Academy already lists the format, the price and eight weighted domains, all about running Claude systems in production at scale. The exam is marked as a preview: there is no registration, no public certification page and no exam guide yet, and Anthropic has announced no launch date. This guide collects everything Anthropic has listed as of November 16, 2026, explains what each domain points to in Anthropic's own documentation, and shows how to start preparing now. We will update it when Anthropic publishes the exam guide.
Status as of November 16, 2026
- Listed: Claude Certified Developer - Professional appears in the Anthropic Partner Academy catalog data, marked as a preview, with its format, price and domains.
- Not open yet: its certification page is not public, it has no checkout and no exam guide, and its prep course is marked as coming soon.
- Not announced: Anthropic's certification FAQ still lists four certifications, and Pearson VUE lists four exams. No launch date has been published.
- Subject to change: the details below are what Anthropic has listed so far, and a preview entry can change before launch.
Exam Quick Facts
What Developer - Professional certifies
Anthropic describes the exam in one sentence: it "validates running Claude-based systems in production at scale: instrumenting them with observability and monitoring, diagnosing complex failures and leading incident response, gating releases with evaluation, designing resilient integrations, controlling cost and latency across a fleet, hardening tool-using agents with layered guardrails, orchestrating durable multi-agent systems, and operating fleets of coding agents for software delivery."
Read that against the Foundations description, which "validates hands-on building with Claude," and the split is clear. Developer - Foundations certifies the engineer who can build a Claude feature correctly: call the API, add tools and MCP servers, pick a model, engineer the prompt. Developer - Professional certifies the engineer who keeps that feature alive after launch, when it serves real traffic, costs real money, breaks at 2 a.m. and gets changed every week by people and by coding agents.
That makes it the first Claude certification centered on operations. Every verb in Anthropic's sentence is something you do to a running system: instrument, diagnose, gate, control, harden, operate.
Try 15 CCDV-P practice questions free
They come from the CCDV-P practice tests, and each option shows its explanation once you answer.
- 6 timed practice exams
- Every answer explained
- A score for each domain
- $19.99 one-time
Exam format
Developer - Professional next to the two exams closest to it (as listed by Anthropic, November 2026)
| Developer - Foundations (CCDV-F) | Developer - Professional | Architect - Professional (CCAR-P) | |
|---|---|---|---|
| Status | Live | Preview (no registration yet) | Live |
| Questions | 53, all scored | 70 (60 scored, 10 unscored) | 63, all scored |
| Time | 120 minutes | 135 minutes | 120 minutes |
| Price | $125 | $175 | $175 |
| Passing score | 720 (scaled 100 to 1,000) | 720 (scaled 100 to 1,000) | 720 (scaled 100 to 1,000) |
| Languages | English and Japanese | English (as listed) | English and Japanese |
| Question types | Multiple choice and multiple response | Multiple choice and multiple response | Multiple choice and multiple response |
| Validity | 12 months | 12 months | 12 months |
| Exam guide | Published | Not yet | Published |
Three details stand out.
Ten unscored questions. The four live Claude exams score every question. Developer - Professional lists 70 questions with 60 scored, so 10 are unscored items, a common way for exam programs to try out new questions on real candidates. You will not know which ten they are, so every question deserves a full answer.
Fifteen more minutes. At 135 minutes for 70 questions, you get about 1.9 minutes per question, close to the pace of the other Claude exams. The extra time pays for the extra questions; scenario-heavy professional exams tend to have longer stems, so plan your pacing on the assumption that reading takes much of the time.
English only, for now. The four live exams are offered in English and Japanese. The Developer - Professional entry lists English. Anthropic added Japanese to the other exams in October 2026, so a Japanese version may follow, but nothing says so yet.
Anthropic has not given the exam a code. Pearson VUE lists the live exams as CCAO-F, CCAR-F, CCAR-P and CCDV-F, so CCDV-P would fit the pattern, but that is our guess until Pearson lists it.
The eight exam domains
Developer - Professional exam domains and weights (as listed by Anthropic)
| Domain | Weight | What it centers on |
|---|---|---|
| Production Integration, Reliability, and Knowledge Grounding | 20% | Keeping Claude calls healthy under real traffic, and grounding answers in your own data |
| Advanced Agent Architecture and Orchestration | 17% | Multi-agent systems that survive long runs, failures and restarts |
| Observability and Production Monitoring | 12% | Knowing what every request did, cost and returned |
| Evaluation Engineering and Quality Assurance | 12% | Eval suites that decide whether a change ships |
| Runtime Guardrails and Production Security | 11% | Layered defenses for agents that call tools |
| Advanced Debugging and Incident Response | 10% | Finding the cause of a production failure and leading the response |
| Agent Fleet Operations for Software Delivery | 10% | Running many coding agents in CI and delivery pipelines |
| Cost and Latency Optimization at Scale | 8% | Controlling spend and response time across a fleet |
The weights tell you where the exam spends its questions. Integration and reliability plus agent architecture make up 37%. The four operations domains (observability, evaluation, guardrails, debugging) together make up 45%, which is the clearest signal that this exam tests running systems more than building them. Our domain-by-domain breakdown maps each domain to the Claude platform features it most likely draws on, with a worked scenario for each.
Production Integration, Reliability, and Knowledge Grounding (20%). The largest domain pairs two jobs. Reliability is about how a Claude integration behaves when the API pushes back: 429 rate-limit errors with a retry-after header, 529 overloaded errors during traffic peaks, long requests that need streaming or the Message Batches API instead of one blocking call, and the request-id on every response that support asks for. Knowledge grounding is about answering from your data: retrieval, citations and search result content blocks.
Advanced Agent Architecture and Orchestration (17%). Anthropic's description says "durable multi-agent systems." Expect orchestrator and subagent designs, context that outgrows the window (compaction, context editing, the memory tool), sessions that resume after a crash, and the choice between a fixed workflow and an autonomous loop at scale.
Observability and Production Monitoring (12%). Every Messages API response reports its usage, including cache_read_input_tokens and cache_creation_input_tokens, and every response carries a request ID. Claude Code exports OpenTelemetry metrics such as claude_code.cost.usage and claude_code.token.usage once CLAUDE_CODE_ENABLE_TELEMETRY=1 is set. This domain is about turning those signals into dashboards and alerts.
Evaluation Engineering and Quality Assurance (12%). "Gating releases with evaluation" means an eval suite runs before a prompt, model or tool change ships, and a regression blocks the release. Anthropic's guidance covers specific, measurable success criteria and three grading methods: code-based, LLM-based (graded by a different model than the one that produced the output) and human.
Runtime Guardrails and Production Security (11%). Tool-using agents read content they did not write: web pages, documents, tool results. "Layered guardrails" means no single defense carries the load: least-privilege tools, approval steps before irreversible actions, schema-valid tool inputs with strict tool use, screening, and handling of the refusal stop reason.
Advanced Debugging and Incident Response (10%). Expect symptom-first scenarios: answers that stop mid-sentence (max_tokens), server-tool loops that pause (pause_turn), empty end_turn responses caused by text added after tool results, a cost spike after a deploy, and quality regressions after a model change. Then the response: contain, roll back, communicate, fix the cause.
Agent Fleet Operations for Software Delivery (10%). Coding agents running unattended: claude -p in CI, the --bare mode Anthropic recommends for scripted runs, permission modes and allowed-tool rules, spend caps, organization-wide managed settings, and agents working in parallel.
Cost and Latency Optimization at Scale (8%). Prompt caching, the 50% discount of the Message Batches API, routing simple traffic to smaller models, effort settings, token counting, and streaming for perceived latency, applied across a fleet where a small inefficiency multiplies.
Three questions at Developer - Professional level
Anthropic has published no sample questions for this exam yet. We wrote these three from the domain names and Anthropic's description, in the scenario style the other Claude exams use, to show the level: each one starts from a running system with a problem.
Three Developer - Professional style questions
A release added a new tool to a customer-support agent. To help the model reason about deadlines, the tool's description now includes the current date and time, filled in when each request is built. Traffic is unchanged, but the Claude bill rose sharply overnight and latency went up. Usage logs show cache_read_input_tokens near zero and cache_creation_input_tokens high on almost every request. What is the most likely cause?
The cache follows the order tools, then system, then messages, and a hit needs an identical prefix. Any change to tool definitions invalidates all of it. A value that changes per request inside a tool description means every request writes a new cache entry (billed above the base input price) and none reads one. Move the timestamp into the latest user message, after the cache breakpoint. Adding a tool makes the prompt longer, so B is backwards, and with steady traffic each read refreshes a 5-minute entry, so C would not explain writes on almost every request.
A research agent uses Anthropic's server-side web search tool. Some answers come back cut short, and the logs show stop_reason: "pause_turn" on the last response of those conversations. The client code treats any stop reason other than tool_use as a final answer and returns it to the user. What should the client do when it sees pause_turn?
pause_turn means a server-tool sampling loop reached its iteration limit, so the turn is unfinished. The fix is to append the assistant content as it is and send the request again. A describes max_tokens, C describes refusal, and server tools run on Anthropic's side, so the client never supplies their results.
A platform team runs Claude Code on every pull request to triage failing tests. They need each run to behave the same on every CI runner, never to wait on a permission prompt, to stop at a spend cap, and to report the cost of each run to their dashboard. Which setup fits all four requirements?
Bare mode skips the hooks, plugins, MCP servers and CLAUDE.md found on the machine, so every runner behaves the same. dontAsk denies anything that would prompt, while the listed tools still run. --max-budget-usd caps spend, and JSON output carries total_cost_usd. Accept-edits mode does not approve arbitrary shell commands such as a test runner, which rules out B.
Full CCDV-P set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsHow it relates to Developer - Foundations (CCDV-F)
Developer - Professional reads as the second half of the developer track. CCDV-F spends a third of its questions on Applications and Integration and only 2.6% on Eval, Testing, and Debugging and 3.1% on Claude Code. Developer - Professional turns those small corners into whole domains: evaluation gets 12%, debugging and incident response 10%, and coding agents in delivery pipelines 10%.
Anthropic lists no formal prerequisite so far, and its FAQ says the Architect - Professional exam has none either, so you may be able to sit Developer - Professional without holding CCDV-F. The content still assumes CCDV-F knowledge throughout: you cannot debug a tool loop you could not build. Our CCDV-F vs Developer - Professional comparison covers which to take first.
How it relates to Architect - Professional (CCAR-P)
Both professional exams cost $175 and both cover integration, evaluation and governance. Architect - Professional is about designing and governing enterprise solutions and leading stakeholders across the lifecycle. Developer - Professional is about operating what was built: monitoring it, debugging it, securing it at runtime and running coding agents. An architect decides which guardrails a system needs, and a developer at professional level implements them and answers the page when one fails. The Developer - Professional vs Architect - Professional comparison goes domain by domain.
Who should take it
- Engineers who own a Claude feature in production, with on-call duty, a cost budget and users who notice when it breaks.
- Platform and infrastructure engineers who run Claude Code or the Agent SDK for many developers and need the organization-level controls.
- Consultants at Claude Partner Network firms who hand systems over to clients and need to show those systems can be run, measured and secured. Like the other Claude exams, it will very likely be open only to people at partner organizations.
- CCDV-F holders looking for the next step on the developer track.
If you have not shipped a Claude integration yet, start with CCDV-F. Its content is the foundation this exam builds on, and it is open for registration today.
How to prepare before the exam guide is out
The exam guide will list the task statements behind each domain, and it is the document to study from once it exists. Until then, the domain names point at skills you can build now with Anthropic's public documentation, and none of this work will be wasted:
- Harden one integration. Take a Claude feature you run and make it survive bad days: retries with backoff that honor
retry-after, handling for 529 overloaded errors, streaming or batches for long outputs, and logging of therequest-idon every call. Handle every stop reason explicitly, includingmax_tokens,refusal,pause_turnandmodel_context_window_exceeded. - Instrument it. Log input, output and cache tokens per request, compute a cache hit rate and a cost per request, and alert when either moves. If your team uses Claude Code, turn on its OpenTelemetry export and build one dashboard from it.
- Build an eval gate. Write success criteria you can measure, build a test set that includes edge cases, grade with code where you can and with a different model where you cannot, and make a failing eval block a merge.
- Attack your own agent. Plant instructions in a document or tool result and see whether your agent follows them. Then add layers: least-privilege tools, an approval step before anything irreversible, strict tool use for schema-valid inputs.
- Run an incident drill. Break something on purpose (a tool description that changes per request, a prompt edit that kills the cache, a model alias that moves) and practice finding it from logs and metrics alone.
- Run a coding agent unattended. Put
claude --bare -pin a CI job with explicit allowed tools, a permission mode, a spend cap and JSON output, and review what it did on ten pull requests.
For the foundations underneath, the CCDV-F complete guide and the CCDV-F practice tests cover the API, agents, tools and MCP knowledge that every Developer - Professional scenario assumes. For the professional level itself, our Early Access Developer - Professional practice exams turn the eight domains into full-length timed exams: 70 questions in 135 minutes, with an explanation for every answer option.
What we don't know yet
- The launch date. Anthropic has published none.
- The exam code. CCDV-P fits the pattern, but nothing official says so.
- The task statements. The exam guide will define what each domain covers, and the catalog entry links to no guide yet.
- Partner Network credit. Developer - Foundations, Architect - Foundations and Architect - Professional count toward an organization's standing in the Claude Partner Network. Whether Developer - Professional will is not stated.
- A Japanese version. The entry lists English.
- Prep content. The entry points to the same free Anthropic Academy courses as Developer - Foundations (the Claude API, Claude Code and an introduction to MCP) and marks a dedicated prep course as coming soon.
Frequently asked questions
The short version
Developer - Professional is Anthropic's certification for the engineer who runs Claude in production: integrations that stay up, agents that survive long runs, metrics that show what happened, evals that gate releases, guardrails that hold under attack, and coding agents running unattended. It is listed at 70 questions in 135 minutes for $175, it is not open yet, and its exam guide will decide the details. The skills behind its domains are public today, so you can start on them before the guide arrives.
Sources:
- Anthropic Partner Academy: Claude certifications catalog
- Claude certification FAQ, Anthropic Partner Academy
- Pearson VUE: Claude Certification Program by Anthropic
- Claude API errors
- Handling stop reasons
- Prompt caching
- Batch processing
- Define success and build evaluations
- Claude Code: run programmatically
- Claude Code: monitoring usage with OpenTelemetry