Skip to content
claude-certified-architect

Refresh history

Changelog

Refresh history for this study guide. Each entry records what changed on Anthropic's side and what was updated here in response.

The exam blueprint itself (domains, weights, scenarios) comes from the official Exam Guide and changes rarely; most entries below are about the platform the exam tests — models, API surface, and Claude Code.

2026-10-14 — Claude Sonnet 5.5 and Haiku 5.5, MCP stateless protocol corrections, Managed Agents workflows

What changed at Anthropic

  • Claude Sonnet 5.5 (claude-sonnet-5-5, released 2026-09-28) and Claude Haiku 5.5 (claude-haiku-5-5, 2026-10-07) complete the 5.5 generation; Sonnet 5 and Haiku 4.5 join the legacy list, and Sonnet 4.5 is deprecated (notice 2026-09-30, retirement 2026-11-30, replacement Sonnet 5.5). Opus 5.5 stays the docs' default for most workloads. Sonnet 5.5 keeps Sonnet 5's $2 / $10 with cache reads at $0.10 (5%), 1M context, 128k output, floor 2027-09-28, and breaks Sonnet 5 code in five ways: thinking: {"type": "disabled"} returns 400 (send {"type": "between_tools"}, accepted at low–high effort, which also fixes effort for the conversation), forced tool_choice returns 400, thinking blocks are bound to the model and the conversation, computer_20251124 is rejected on the Claude API and Google Cloud, and some advisor pairings return 400 (advice from 5.x advisors arrives encrypted). Its effort levels are recalibrated (API default high, but start agentic coding at medium), text between tool calls arrives as thinking blocks, and it gains mid-conversation system messages, per-message effort, task budgets, and a 512-token cache minimum. Haiku 5.5 is 1M / 128k with prompt-length pricing — $0.10 / $0.50 up to a 100k-token prompt, $0.50 / $2.50 above (the only 4.6-or-later model without flat 1M pricing) — default effort medium, disabled still allowed at high or below, forced tool_choice still accepted (with no thinking block), and breaks Haiku 4.5 code on budget_tokens, sampling parameters, prefill, computer_20250124, Bedrock structured outputs, and edited history. Sonnet 5.5 and Haiku 5.5 thinking blocks are also account-bound (organization_binding_mismatch), both run safety classifiers, and Haiku 5.5 has no server-side refusal fallback.
  • Feature coverage widened to the new models: compaction on demand, inline tools and mid-conversation tool changes, mid-conversation system messages (with a new caveat — Sonnet 5.5 / Haiku 5.5 may ignore a system message placed right after a tool_result), per-message effort, preserved-thinking enforcement (now Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5), task budgets, tool search, and the 300k batch output beta. Computer use on Claude API / Google Cloud is toolset-only for every 5.5 model. The Models API gained line, lifecycle fields, capabilities.thinking.types, and capabilities.server_tools; the Admin API went GA in the SDKs (client.organization, ant organization).
  • Managed Agents: dynamic workflows and subagents via multiagent: {"type": "multiagent_20261001"} (2026-10-09; ant v1.40.0), with server-run workflow programs, workflow_run.* events, and documented limits; web_fetch now only fetches URLs already seen in the session, and limited networking applies allowed_hosts to the web tools (2026-10-07); ant apply manages vaults (v1.34.0).
  • SDKs: the Tool Runner can run tools while the reply streams (run_tools_eagerly, defer_tool_call); beta browser and computer toolset classes (2026-10-07); client-side compaction_control is removed (TypeScript 0.129.0, Python 1.0.0). Agent SDK (TypeScript 0.3.286): omitting permissionMode no longer means default.
  • MCP: no new protocol revision — 2026-07-28 is still current. SEP-2127 (MCP Server Cards) was finalised as an extension.
  • Claude Code v2.1.282 → v2.1.299: terminal and VS Code sessions with no mode configured now start in auto mode (v2.1.283; -p and the Agent SDK still start in default except on third-party providers or with telemetry off); sonnet and haiku now resolve to Sonnet 5.5 (v2.1.284) and Haiku 5.5 (v2.1.293), and all three 5.5 models start at medium effort; ultracode became its own toggle instead of an effort level (v2.1.284) and the Agent tool takes a per-call effort (v2.1.292); hooks gained onFailure: "block" to fail closed (v2.1.295), and matcher or serialization failures on PreToolUse / PermissionRequest now block instead of being skipped (v2.1.288); mods — plugins that ship in-process JS/TS functions — arrived with allowManagedModsOnly (v2.1.287); nested CLAUDE.md and path-scoped rules now load on Write / Edit and single-file Bash reads (v2.1.288 / v2.1.293); the AGENTS.md plugin was renamed and "Project instructions" gained four values (v2.1.285); MCP description caps rose from 2,048 to 4,096 characters (16,384 for tool-search-loaded tools; first-party servers uncapped), alwaysLoad: false defers every tool, mcp__a rules match only server a, and 2026-07-28 protocol negotiation spread to stdio, claude.ai, and cloud connectors; --bare now loads only command-line MCP servers and runs no background tasks (v2.1.286); idle compaction (v2.1.299), /doctor prompt-audit, /verify, pre-commit verify / simplify skills, claude purge, --desktop, and managed availableModelsMatch / deniedModels / allowedProviders. The hook catalog is unchanged at 33 events.
  • Certification program: no new exam guide version or credential. The live policies page puts the free cancel / reschedule window at 48 hours (the v1.0 guide said 24); renewal is an open-book assessment you can retake; the FAQ lists the exams in English and Japanese, and Japanese translations of all four v1.0 exam guides were published on 2026-10-08; CCAO-F does not count towards Claude Partner Network tier eligibility. Anthropic also announced Claude Frontier Academy (2026-10-02), a nominated residency that awards badges, not a new proctored exam.

What changed here

  • README and main guide: lineup table now Fable 5.1 / Opus 5.5 / Sonnet 5.5 / Haiku 5.5, with Sonnet 5 and Haiku 4.5 moved to legacy and Sonnet 4.5's retirement; the "what the 5.x generation changes" list rewritten for four models (who rejects forced tool_choice, the new thinking-read matrix, account-bound blocks, append-only history on all four); Sonnet 5.5's between_tools and Haiku 5.5's breaking changes; retirement floors and the 60-day notice; Mythos 5.1 access wording updated to Anthropic's verification programs; decision-framework rows for tool_choice, thinking, effort, agent harness (self-hosted Managed Agents sandboxes), history, progress UX, and orchestration; program facts and new resource links.
  • Samples: the forced-tool_choice samples move from claude-sonnet-5 (now legacy) to claude-haiku-5-5, the only current model that still accepts forced tool use — with a note that a forced call skips thinking.
  • Domain 1: refusal classifiers on Sonnet 5.5 / Haiku 5.5, no server-side fallback (or fallback credit) on Haiku 5.5, Sonnet 5.5's fallback routing and the between_tools fallback rule; the thinking-read matrix as a table, plus account binding; task budgets on the new models; Tool Runner eager tool execution and defer_tool_call(); Managed Agents self-hosted sandboxes, vaults, multiagent_20261001 (subagents, workflow runs, advisor, inline agents, limits, idempotent tools), and the web_fetch prior-context rule; per-call effort on the Agent tool; ultracode as a toggle; Agent SDK permissionMode default and autoCompactWindow; hooks onFailure: "block", fail-closed matching, checks on rewritten input, and UserPromptSubmit before skill expansion (changelog-sourced).
  • Domain 2: forced-tool-use box, sample, and Q5 cover Sonnet 5.5 and Haiku 5.5; tool search model coverage; advisor pairings and encrypted advisor_redacted_result; computer use moved to computer_toolset_20260801 on every 5.5 model, plus the SDK browser / computer toolset classes; mid-conversation tool changes on the new models, header scope per platform, and the after-tool_result placement caveat; MCP description limits as a table, alwaysLoad: false, startup-wait exception, widgets reserved name, /mcp reconnect all, exact mcp__ rule matching, and wider 2026-07-28 negotiation.
  • Domain 3: auto as the starting permission mode and "Manual" for default; lazy CLAUDE.md / rules loading on writes and Bash reads; AGENTS.md "Project instructions" values; /doctor prompt-audit, /verify, /run-skill-generator, pre-commit verify / simplify, forked skills inside workflows; a mods section with a mod / settings hook / skill / MCP comparison; plugin component list and --config; ultracode toggle, per-call effort, 32-skill preload cap; --bare limits, unattended background-command timeouts, combinable system-prompt flags, new env vars, claude purge, --desktop; 5.5 model aliases and effort defaults; managed model and provider restrictions; /autocompact per model and idle compaction; batch 300k list.
  • Domain 4: forced tool_choice and batch sample on Haiku 5.5; 300k batch list; Sonnet 5.5 and Haiku 5.5 rows in the thinking table (between_tools, disabled rules, sampling parameters) with Sonnet 5 / Haiku 4.5 marked legacy; effort defaults and per-model starting points, per-message effort on the new models; display defaults and progress notes; prefill; mid-conversation system messages with the placement caveat; preserved thinking rewritten (read matrix, enforcement on all four, account binding, block_binding only with adaptive); caching minimums, cache-read rates, Haiku 5.5's tiered pricing.
  • Domain 5: context table split into current and legacy; tokenizer note for the new models; Haiku 5.5 traps (5x price above a 100k-token prompt, every earlier turn's thinking kept in context, no context-awareness tags); idle compaction; kept-turn thinking, context editing, and on-demand compaction on all four 5.5-generation models; task budgets on every current model; refusals on Sonnet 5.5 / Haiku 5.5.
  • certs/: 48-hour reschedule window, open-book renewal, English / Japanese delivery, and (CCAO-F) no Partner Network tier credit.
  • Moved links updated: Building Effective Agents is now under anthropic.com/engineering, the MCP introduction lives at modelcontextprotocol.io/docs/getting-started/intro, and the API docs root is platform.claude.com/docs/en/intro.

Corrections of earlier errors

  • MCP lifecycle (Domain 2): the guide described the 2026-07-28 revision as adding URL-mode elicitation. URL mode dates from 2025-11-25; 2026-07-28 is the stateless revision — no initialize handshake or Mcp-Session-Id, a required server/discover, version and capabilities in _meta, Multi Round-Trip Requests (resultType: "input_required"), and Roots, Sampling, and Logging deprecated. The handshake is now labelled as the pre-2026-07-28 lifecycle, and notifications/initialized is attributed to the client.
  • Domain 1: Setup does not run once per session — it fires only with --init-only, or --init / --maintenance in -p mode; permissionDecision was missing "defer" and the deny > defer > ask > allow precedence; "Fable 5.1 reads everything" no longer holds now that it can't read Sonnet 5.5 or Haiku 5.5 blocks.
  • Domain 3: CLAUDE.md files are concatenated, not overridden; @ imports go four hops deep, not five; subagent-output scanning still runs alongside the v2.1.277 header; an allowed-tools grant clears at your next message; dontAsk denies anything that would prompt.
  • Domain 5: after /compact, only the project-root CLAUDE.md and unscoped rules are re-injected — nested CLAUDE.md files and path-scoped rules reload when Claude next touches a matching file; client-side SDK compaction_control is removed, not deprecated (the platform docs still say "deprecated").
  • Main guide: the lineup said Haiku 4.5's floor was the only one inside the next year; Opus 4.5's (2026-11-24) was nearer, and Sonnet 4.5 is now scheduled to retire.

Deliberately left out

  • Claude Code changelog-only items with no docs page yet (subagent-frontmatter autoCompactWindow, CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL, claude plugin eval --disallow-tools), the refusal stop reason on Managed Agents idle events (SDK types only), and whether Japanese exams can already be booked at Pearson VUE.
  • Haiku 5.5's exact refusal category list (only Sonnet 5.5's is documented), and roughly 1,300 cosmetic Claude Code changelog entries (VS Code, Slack, gateway, UI polish).

2026-09-25 — Claude Opus 5.5, compaction on demand, inline tools

What changed at Anthropic

  • Claude Opus 5.5 (claude-opus-5-5, released 2026-09-22) is now the docs' default for most workloads; Opus 5 joins the legacy list. $4 / $20 per MTok with cache reads at $0.20 (5% of input, vs 10% on most models and 2.5% on Fable 5.1), 1M context, 128k output, retirement floor 2027-09-22. It inherits Fable 5.1's three breaking changes (no forced tool_choice, model-bound thinking blocks, append-only history) and adds its own: thinking can't be disabled, the default effort is medium (Opus 5 was high), and on the Claude API and Google Cloud computer use works only through computer_toolset_20260801. Without any error, the text Claude writes between tool calls now arrives as thinking blocks, and a biology classifier plus a reasoning_extraction category join cyber. Fast mode is available on it ($8 / $40). Priority Tier is not, and new Priority Tier commitments are no longer sold. Sonnet 5's $2 / $10 introductory price is now permanent.
  • Compaction on demand (beta compact-2026-09-04, 2026-09-14): send the conversation with a top-level compaction: {"type": "summarize"} and the API returns a signed compaction block (stop_reason: "compaction") that you then send first, in place of the messages it summarized. You choose the moment, can keep recent turns word for word, can run it in the background, and can supply your own instructions. On Fable 5.1 / Opus 5.5 the thinking in kept turns stays valid when the API wrote the summary. The docs now say to prefer it over threshold compaction. Not on Bedrock.
  • Inline tools (beta inline-tools-2026-09-15, 2026-09-22): a tool_addition block in a mid-conversation system message can carry a full tool_definition, so you can add a tool, change its schema, or move a server tool to a newer version without editing tools or losing the cache; with mcp-client-2026-09-15 the definition can be an MCP toolset, and responses record fetched tool lists in an mcp_tool_listing block.
  • Preserved thinking: enforcement covers Fable 5.1 and Opus 5.5; Opus 5.5 reads Opus / Sonnet / Haiku blocks but not Fable / Mythos ones, and only Fable 5.1 / Mythos 5.1 read Opus 5.5's. Under thinking-binding-controls-2026-08-01, input_transformations gained thinking_mismatch_allowed entries (2026-09-14) that flag history edits on accounts where the check isn't enforced yet.
  • Cache diagnostics is GA (2026-09-23, no beta header); the diagnostics field is always present, and a fingerprint is stored only when the request includes the diagnostics object.
  • Refusal billing (2026-09-24): a refusal that arrives before any output is billed when its category is bio, frontier_llm, or reasoning_extraction; general_harms joined the documented categories.
  • Managed Agents / ant CLI: ant apply manages agents, environments, skills, memory stores, and deployments as files with a claude-lock.json lockfile (v1.30.0, 2026-09-03); permission policies gained auto (server evaluates each tool call) and ant beta:sessions connect attaches a terminal to a live session (v1.32.0, 2026-09-10). Per-message effort also reached Google Cloud.
  • Claude Code v2.1.263 → v2.1.282: Opus 5.5 is the default Opus model and Pro / Team Standard plans default to Opus; effort is per model with a maxEffortLevel cap; AGENTS.md is read when a project has no CLAUDE.md (v2.1.277, every provider since v2.1.281); auto mode's classifier runs server-side by default; subagent results reach the main agent framed as subagent output; the medium workflow-size guideline dropped from 15 to 10 agents (Pro defaults to small) and CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS raises concurrency; omitClaudeMd agent frontmatter; PermissionRequest no longer runs agent-type hooks; the TaskOutput tool was removed (taskOutputMaxChars is inert); /output-style, --system-prompt-snapshot off, "attribution": false, CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH, CLAUDE_CODE_MCP_STARTUP_WAIT_MS, MCP 2026-07-28 negotiation everywhere, claude plugin validate MCP checks, --plugin-dir folders, claude.ai skill and plugin sync; and a run of fixes so --continue / --resume stop re-sending history in a changed form (the preserved-thinking check biting Claude Code itself).
  • Certification program: the Partner Academy FAQ now states that delivery moved to Pearson VUE and badging to Credly on 2026-06-30, that partner-tier discounts apply (50% for Select / Preferred / Global Premier, 100% for Global Premier through 2026-12-31), that candidates must be 18+, and that items are multiple-choice and multiple-response.

What changed here

  • README: on request, the website and PDF links moved into the intro and the "Read it online" section was removed. Lineup led by Opus 5.5; code samples moved from claude-opus-5 to claude-opus-5-5, except the forced-tool_choice samples, which now use claude-sonnet-5 with a note.
  • Main guide: lineup table, Opus 5.5 breaking and behaviour changes, Priority Tier note, decision-framework rows (effort default per model, tool_choice, history, on-demand compaction, progress UX), program facts, resource links.
  • Correction: the September 7 refresh said Mythos 5.1 runs safety classifiers. The docs don't say that — it is the same model served without Fable 5.1's dual-use measures, with safeguards that depend on the access program. Reworded in the main guide, Domain 1, and Domain 5.
  • Domain 1: refusal categories and billing; server-side fallback on Opus 5.5 and the thinking-read matrix; task budgets on Opus 5.5; ant apply, the auto permission policy, and sessions connect; workflow sizing; PermissionRequest / mcp_tool hook notes (hook catalog re-verified at 33 events).
  • Domain 2: forced-tool-use box and Q5 cover Opus 5.5; new "Defining tools in a message" material under §2.6; MCP description cap, startup wait, and protocol negotiation notes.
  • Domain 3: AGENTS.md fallback rewritten; server-side auto-mode classifier; omitClaudeMd; subagent output framing; bundled skills (/fast on Opus 5.5, /output-style, /insights, new /claude-api subcommands); plugin tooling; settings table (maxEffortLevel, attribution, sync flags, per-model effort; taskOutputMaxChars dropped); batch 300k on Opus 5.5; Remote Control forks.
  • Domain 4: Opus 5.5 row in the thinking table and the can't-disable rule; effort guidance rewritten around the medium default; per-message effort on Opus 5.5; progress notes as thinking blocks on Opus 5.5; prefill and mid-conversation system messages on Opus 5.5; preserved-thinking section rewritten (model matrix, enforcement scope, thinking_mismatch_allowed); inline tools pointer; caching (Opus 5.5 512-token minimum and 5% reads, cache diagnostics GA).
  • Domain 5: context table and tokenizer note; compaction section rewritten as on-demand vs threshold; "shapes that stay valid" table updated for API-written keep-tail and background compaction; context editing never invalidates thinking; choosing table; task budgets on Opus 5.5 (and the remaining / compaction conflict); refusal billing; new practice question Q8.
  • certs/: Credly badging, partner discounts, and the age requirement added to each overview.

2026-09-07 — Claude Fable 5.1, preserved thinking, dynamic workflows

What changed at Anthropic

  • Claude Fable 5.1 (claude-fable-5-1, released 2026-09-01) is the most capable widely released model; Claude Mythos 5.1 is its Project Glasswing twin and, unlike Mythos 5, runs safety classifiers. Fable 5 joins Opus 4.8/4.7/4.6/4.5 and Sonnet 4.6/4.5 on the legacy list; Opus 4.1 retired 2026-08-05. Same $10/$50 pricing as Fable 5, cache reads down to $0.25/MTok. Anthropic now publishes per-model retirement floors.
  • Three breaking changes on Fable 5.1: forced tool_choice (any / tool) returns 400; thinking blocks are readable only by the producing model or a newer one (a fallback to an older model drops them, unbilled); and editing earlier turns invalidates later thinking blocks — harnesses must keep history append-only or trim on the server. New organizations (created on or after 2026-08-31) already get a 400 on edited history.
  • Additive on Fable 5.1 / Mythos 5.1 (beta): per-message effort inside a role: "system" message (also on Opus 5), turn-scoped system messages (clear_at: "next_user_message"), thinking.display: "updates" progress notes, tool_addition / tool_removal blocks, and content provenance.
  • Task budgets are not supported on Sonnet 5 (the docs' feature table); a budget that is too small produces refusal-like scoping-down. Prompt-cache minimums are model-dependent (512 tokens on Opus 5 / Fable 5.x, up to 4,096 on Opus 4.6 and Haiku 4.5). Effort guidance changed: start at high and sweep on Opus 5 / Fable 5.1; xhigh remains the starting point on Opus 4.8/4.7.
  • The Admin API landed in every SDK (client.beta.organization) and the ant CLI (ant auth login, ant beta:agents create < agent.yaml) is now the recommended control plane for Managed Agents. The Python SDK crossed to 1.x.
  • Claude Code (v2.1.236 → v2.1.263): dynamic workflows (Workflow tool, /workflows, /workflow-authoring, /deep-research, ultracode) and experimental agent teams; PreModelSwitch / PostModelSwitch hooks (33 events); subagent_type: "fork" with fork mode on by default, /fork now copying a session into a background session and /branch for in-place branching; background-session lifecycle (claude --bg, claude agents, attach / logs / stop / respawn / rm); --restricted, --permission-prompts none, --safe-mode, --append-subagent-system-prompt-file; /skill-doctor, /diff, /advisor, /fast, /feedback; managedMcpServers, promptCacheTtl, modelPicker, modelPricing, output-cap settings; auto memory as one typed file per memory behind a MEMORY.md index.
  • Certification program: the Partner Academy FAQ confirms free on-time renewal via a non-proctored assessment (a lapsed credential means the full exam again) and partner-domain-email registration. The v1.0 CCAR-F Exam Guide PDF is now linked directly.

What changed here

  • README and main guide: lineup table led by Fable 5.1 with the three breaking changes summarised; retirement floors; renewal policy; official v1.0 exam guide and Partner Academy links; decision-framework rows for tool_choice (per-model), effort (start at high), append-only history, and subagents vs workflows; models-docs URL updated to platform.claude.com/docs/en/models/overview.
  • Domain 1: fallback consequences on Fable 5.1; task-budget availability corrected (no Sonnet 5) and too-small-budget behaviour; ant CLI flow for Managed Agents; a subagents / skills / agent teams / workflows comparison; subagent_type: "fork", /fork vs /subtask vs /branch; hook catalog 31 → 33 (PreModelSwitch, PostModelSwitch); background-session commands.
  • Domain 2: Fable 5.1 forced-tool-use box and replacement patterns; managedMcpServers scope; ToolSearch, SendMessage / ListAgents, and Workflow in the built-in tools table; new section on tool_addition / tool_removal; new practice question Q5.
  • Domain 3: background sessions, /fork, /loop self-pacing; typed auto-memory files; refreshed bundled-skill list; experimental.cacheTtl; new "Agent Teams and Dynamic Workflows" section; new CI flags (--permission-prompts none, --restricted, --safe-mode, --append-subagent-system-prompt-file); new cost/cache/fleet settings table.
  • Domain 4: Fable 5.1 row in the thinking table; rewritten effort guidance with per-message effort; display: "updates"; turn-scoped and tool-change system messages; new "Preserved Thinking" section; caching table with per-model minimums, pricing, and cache diagnostics; extra silent invalidators.
  • Domain 5: Fable 5.1 in the context table and tokenizer note; compaction availability and instructions; new "Compaction Shapes That Stay Valid" section; task-budget corrections; Mythos 5.1 refusals; new practice question Q7.
  • certs/: renewal policy added to each overview.

2026-08 — Claude Opus 5, tool search, server-side context management

What changed at Anthropic

  • Claude Opus 5 (claude-opus-5) is the current default for complex agentic coding and enterprise work. Opus 4.8, 4.7, 4.6, Sonnet 4.6, Sonnet 4.5, and Opus 4.5 moved to the legacy list. Claude Mythos 5 joined Fable 5 under Project Glasswing (invitation-only). Sonnet 5 lists at $2 / $10 per MTok.
  • Tool search is GA on the Claude API and on by default in Claude Code. Anthropic's published threshold for tool-selection degradation is now ~30–50 tools, with defer_loading as the scaling mechanism.
  • Server-side context management matured: context editing (clear_tool_uses_20250919, clear_thinking_20251015), compaction (compact_20260112), and the memory tool (memory_20250818).
  • Task budgets (beta) give an agentic loop a ceiling the model can see.
  • Server-side refusal fallbacks (fallbacks: "default") route around a refusal stop reason automatically.
  • Claude Code shipped as a desktop app and on the web/mobile, alongside the terminal, VS Code, and JetBrains; sessions move between surfaces (--cloud, --teleport, /desktop, Remote Control). Routines run scheduled work in the cloud.
  • Claude Code's subagent tool is canonically Agent (formerly Task), with custom subagents defined in .claude/agents/*.md.

What changed here

  • All code samples moved from claude-opus-4-8 to claude-opus-5; added a current model lineup table to the main guide and README.
  • Domain 1: added server-side refusal fallbacks; a task-budget note under the iteration-cap anti-pattern; a "who runs the loop" comparison (manual loop / Tool Runner / Managed Agents / Agent SDK) including Managed Agents; corrected the Task → Agent tool naming; forks vs fresh subagents; session portability; hook catalog updated to 31 events (added DirectoryAdded).
  • Domain 2: kept the exam's "4-5 tools" answer and added the tool search reality alongside it; modernised MCP configuration (transport type, scopes, claude mcp add, OAuth, alwaysLoad, ENABLE_TOOL_SEARCH, output limits); new §2.6 on Anthropic-defined and server-side tools, parallel tool use, programmatic tool calling, and the MCP connector.
  • Domain 3: new §3.0 on surfaces, session portability, and the three scheduling mechanisms; plugins and claude plugin eval; a full custom-subagent section (.claude/agents/ frontmatter, built-in agents, permission inheritance, nesting and concurrency limits); expanded CI flag table and managed CI integrations; refreshed bundled-skill list.
  • Domain 4: new §4.7 on adaptive thinking, effort, thinking display, the removal of prefill and sampling parameters, and mid-conversation system messages; new §4.8 on prompt caching; output_format marked deprecated; batch extended output (300k) and result ordering.
  • Domain 5: current context-window sizes and the Opus 4.7 tokenizer change; new §5.7 on context editing vs compaction vs memory tool; new §5.8 on task budgets and refusals as a reliability concern.
  • Resources reorganised by source and the stale docs/en/docs/... tool-use link fixed. The generated PDF now includes the certs/ overview guides.

2026-07 — Pearson VUE program update, four credentials

  • Certification family expanded to four credentials (CCAO-F, CCAR-F, CCAR-P, CCDV-F); added overview guides in certs/.
  • Exam delivery moved to Pearson VUE (OnVUE online or test center), registered through the Anthropic Partner Academy. Exam Guide v1.0 (effective July 2026) set a $125 fee for CCAR-F and 12-month validity, superseding the launch-period "$99 / first 5,000 partner employees free" terms.
  • Documentation split: API docs to platform.claude.com, Claude Code docs to code.claude.com.
  • Schema, tool_choice, and hooks corrections throughout.

2026-06 — Opus 4.8 lineup

  • Refreshed the model lineup (Fable 5, Sonnet 5) and added the refusal and model_context_window_exceeded stop reasons.
  • Added the landing page and auto-built PDF via GitHub Pages.

2026-04 — Docs URL migration, Agent SDK rename

  • Migrated documentation URLs; renamed "Claude Code SDK" to Claude Agent SDK; expanded the hooks catalog.