> For the complete documentation index, see [llms.txt](https://docs.agenticflow.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.agenticflow.ai/changelog/office-hour-51-claude-model-updates.md). # Office Hour #51: Claude Fable 5, Sonnet 5 & Opus 4.8 Model Updates ### Quick Recap Office Hour #51 expands model availability across PixelML and the Anthropic provider, giving builders more Claude-family options for high-capability reasoning, long-running agent work, and everyday cost-performance workflows. This release includes: * Claude Fable 5 availability through PixelML * Claude Sonnet 5 availability through PixelML * Claude Opus 4.8 availability through the Anthropic provider * Claude Fable 5 availability through the Anthropic provider * Claude Sonnet 5 availability through the Anthropic provider * Clearer model-choice guidance for builders choosing between frontier capability, enterprise reliability, and everyday speed/cost balance * AgenticFlow CLI v1.10.7 updates, including the new `autonomous-desk` workforce blueprint, a MAS graph-building playbook, workflow-in-workforce routing support, and a much larger CLI-shipped blueprint catalog 🎥 Watch the full session: {% embed url="" %} *** ### What Changed & Why #### 1. New Claude Models Added Across PixelML and Anthropic This release adds new Claude model options to the model catalog so agents and workflows can be configured against the right provider surface for each team. **PixelML Provider** * Claude Fable 5 * Claude Sonnet 5 **Anthropic Provider** * Claude Opus 4.8 * Claude Fable 5 * Claude Sonnet 5 Teams now have a cleaner path for selecting Claude models directly from the Model Selector, whether they prefer PixelML-managed access or their own Anthropic connection. ```mermaid flowchart LR W["Workload"] --> S["Claude Sonnet 5"] W --> O["Claude Opus 4.8"] W --> F["Claude Fable 5"] S --> S1["Everyday production agents"] S --> S2["High-volume coding, document, support, and content work"] O --> O1["Complex enterprise coding"] O --> O2["Longer agent sessions where reliability matters"] F --> F1["Highest-capability reasoning"] F --> F2["Deep research, long-horizon planning, and vision-heavy workflows"] F --> R{"Safety classifier refusal?"} R -->|"No"| Done["Return Fable result"] R -->|"Yes"| FB["Fallback to another Claude model"] ``` The selection path is simple: use Sonnet 5 for scalable everyday execution, Opus 4.8 for dependable high-end enterprise agent work, and Fable 5 when the workflow needs the highest available capability and can handle refusal/fallback behavior. #### 2. Better Coverage Across Capability Tiers These additions give builders a more complete Claude lineup for different workload profiles: * **Claude Fable 5** is the highest-capability option for the most demanding reasoning, long-horizon agentic work, complex software tasks, research, vision-heavy workflows, and multi-step planning. * **Claude Opus 4.8** is a strong default for complex agentic coding and enterprise work where reliability, instruction following, and sustained context handling matter. * **Claude Sonnet 5** is the balanced option for everyday agent workflows, coding, document work, and general knowledge tasks where teams want strong capability with better speed and cost efficiency. For practical agent design, use Fable 5 when maximum capability matters most, Opus 4.8 when a complex enterprise workflow needs a dependable high-end model, and Sonnet 5 when you want the best day-to-day balance. **Model Evidence Notes** The notes below summarize external model evidence from Anthropic's public model documentation and announcements. They are useful for model-selection context, but should be read as publisher-provided model evidence rather than AgenticFlow-run evaluation results. | Model | Surface | Why it matters | | --------------- | ------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Claude Fable 5 | PixelML, Anthropic | Anthropic describes Fable 5 as its most capable widely released model, built for demanding reasoning and long-horizon agentic work. It supports a 1M-token context window, up to 128K output tokens, adaptive thinking, tool calling, code execution, memory, and vision. | | Claude Opus 4.8 | Anthropic | Anthropic positions Opus 4.8 for complex agentic coding and enterprise work. It supports a 1M-token context window, up to 128K output tokens, adaptive thinking, vision, and strong tool-use workflows. | | Claude Sonnet 5 | PixelML, Anthropic | Anthropic positions Sonnet 5 as the best combination of speed and intelligence, with substantially improved cost-performance over Sonnet 4.6 and higher-effort performance that can match Opus 4.8 on some tasks. It supports a 1M-token context window, up to 128K output tokens, adaptive thinking, vision, and tool-use workflows. | **Visual Model Evidence** The figures below are publisher-provided charts from Anthropic. They add visual context for why these models are useful in agent and workflow builders, but they are not AgenticFlow-run benchmark results.
Official Claude Opus 4.8 benchmark table comparing coding, agentic terminal coding, reasoning, computer use, knowledge work, and financial analysis results

Claude Opus 4.8 benchmark evidence from Anthropic's Opus 4.8 announcement. Source: Introducing Claude Opus 4.8.

Official Claude Sonnet 5 agentic search cost-performance chart comparing Sonnet 5, Opus 4.8, and Sonnet 4.6 across effort levels

Sonnet 5 cost-performance evidence on agentic search from Anthropic. Source: Introducing Claude Sonnet 5.

Official Claude Sonnet 5 agentic computer use cost-performance chart comparing Sonnet 5, Opus 4.8, and Sonnet 4.6 across effort levels

Sonnet 5 cost-performance evidence on agentic computer use from Anthropic. Source: Introducing Claude Sonnet 5.

**Builder Integration Notes** Fable 5 is the one model in this group that needs extra integration handling. Anthropic documents that Fable 5 can return `stop_reason: "refusal"` as a successful HTTP 200 response when its safety classifiers decline a request. For production agents, that means builders should treat Fable refusals as a normal model outcome, not as infrastructure failure. Practical defaults: * Add a visible fallback path for Fable 5 workflows, usually to Opus 4.8 for complex work or Sonnet 5 for cost-sensitive everyday execution. * Log refusals separately from provider errors so support teams can distinguish policy behavior from outages. * Use the `effort` setting intentionally on Opus 4.8 and Sonnet 5 instead of relying on defaults when cost, latency, or quality targets matter. * For high-volume Sonnet 5 workloads, account for the launch pricing window and the standard pricing that applies after August 31, 2026. Sources: * [Anthropic model overview](https://platform.claude.com/docs/en/about-claude/models/overview) * [Introducing Claude Fable 5 and Claude Mythos 5](https://platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5) * [Introducing Claude Opus 4.8](https://www.anthropic.com/news/claude-opus-4-8) * [Introducing Claude Sonnet 5](https://www.anthropic.com/news/claude-sonnet-5) #### 3. How to Choose the New Models Use the new model entries as workload-specific defaults: * Choose **Claude Fable 5** for high-stakes planning, deep research, complex code generation, large migrations, multi-step agent loops, and workflows where quality matters more than cost or latency. * Choose **Claude Opus 4.8** for enterprise-grade coding, analysis, document-heavy workflows, and longer agent sessions that need strong consistency and lower failure risk. * Choose **Claude Sonnet 5** for production agents, support workflows, content operations, summarization, coding assistance, document analysis, and other high-volume workflows where capability and cost both matter. #### 4. AgenticFlow CLI v1.10.7: More Blueprints and a Self-Correcting Desk The AgenticFlow CLI has also moved from the v1.10.0 surface documented in earlier Office Hours to `@pixelml/agenticflow-cli@1.10.7`. Key CLI updates since the last changelog pass: * **New `autonomous-desk` workforce blueprint.** This is a desk-topology MAS pattern: Planner → route gate → deterministic workflow or web Researcher → Critic QA gate → automatic revision → Editor. Use it when a mission needs quality control, revision, and a final edited deliverable rather than a one-pass answer. * **Workflow-in-workforce routing.** `af workforce init` now supports `--tool-workflow-id`, `--tool-workflow-purpose`, and `--tool-workflow-input` for the desk topology, so a deployed public-runnable workflow can serve as the deterministic route inside a higher-autonomy workforce. * **New `mas-graph-building` playbook.** This documents the field-verified rules for hand-authored workforce graphs: the required `.output` hop in template references, condition branch indexes, state variable merging, structured-output schemas, and workflow invocation through `call_other_workflow`. * **Larger CLI blueprint catalog.** The CLI now ships **379 blueprints**: 362 workflow blueprints, 3 agent blueprints, and 14 workforce blueprints. That includes n8n migration and many practical workflow packs such as `n8n-converter`, `email-to-structured`, `rss-digest-email`, `competitor-url-snapshot`, and `job-app-package`. * **Smoother agent and workflow execution.** Recent v1.10.x fixes raised blueprint agent recursion defaults to 100, added `af workflow run --wait`, returned canonical workflow/workforce Web UI links in `_links`, added safe invocation guidance in `af bootstrap --json`, and added a Node.js 18+ launcher guard for npm installs. Practical operator guidance: * Start every AI-assisted CLI session with `af bootstrap --json`; the returned `invocation`, `blueprints`, `commands`, `changelog`, and `_links` fields are the current source of truth. * Use `af blueprints list --kind workflow --json` before reaching for a workforce. Most new catalog growth is deterministic workflow packs, not multi-agent teams. * Use `af workforce init --blueprint autonomous-desk --json` for self-correcting research, planning, or review missions; read `af playbook mas-graph-building` before hand-authoring a workforce graph. * Use `af workflow run --workflow-id --body '' --wait --json` when you want a single command to start the workflow and return the final output. *** ### Model Updates Included * Added Claude Fable 5 to the PixelML provider * Added Claude Sonnet 5 to the PixelML provider * Added Claude Opus 4.8 to the Anthropic provider * Added Claude Fable 5 to the Anthropic provider * Added Claude Sonnet 5 to the Anthropic provider * Added model-choice guidance for selecting between Fable 5, Opus 4.8, and Sonnet 5 ### CLI Updates Included * Updated the documented current CLI version to `@pixelml/agenticflow-cli@1.10.7` * Added the `autonomous-desk` workforce blueprint to the Office Hours release record * Documented workflow-in-workforce desk flags: `--tool-workflow-id`, `--tool-workflow-purpose`, and `--tool-workflow-input` * Added the `mas-graph-building` playbook as the source of truth for hand-authored MAS graph rules * Updated blueprint counts to 379 total CLI-shipped blueprints: 362 workflow, 3 agent, and 14 workforce * Added `af workflow run --wait` and canonical `_links` behavior to the CLI release summary *** ### 🚀 5 Workflow Blueprint Prompts Copy any prompt below into **Claude Code**, **Codex**, or another AI coding agent with the **AgenticFlow CLI** installed to deploy a working workflow in minutes. These five use cases cover security triage, market intelligence, stock research, sales-call coaching, and local lead enrichment. #### 1. Web Security Quickscan Packet (`network-vulnerability-report-pack`) Turn raw scan findings into a prioritized security triage report — executive summary, severity-ranked findings table, likely CVE lookup targets, remediation plan, and a notification draft. Draft-only; it does not run scans or take external actions. > Using the AgenticFlow CLI, deploy the network-vulnerability-report-pack blueprint. Use these scan findings: > > Host: api.internal.example | Open ports: 22/tcp OpenSSH 8.2, 80/tcp nginx 1.18, 443/tcp nginx 1.18 | TLS: supports TLS 1.2 and 1.3 | Finding: HTTP redirects to HTTPS but missing HSTS | Finding: SSH password authentication enabled | Scope: internal staging API > > Report policy: Create a security triage report with severity, likely CVE lookup targets, remediation steps, and notification draft. Draft only; do not claim Nmap execution, file write, Telegram send, email send, CVE API lookup, or external action occurred. > > Deploy, run once, show me the executive summary, findings table, CVE lookup targets, remediation plan, notification draft, and assumptions. *** #### 2. Market Intelligence Brief (`market-intelligence-brief-pack`) Turn scattered regional market signals into a decision-ready brief with an executive summary, regional insights, an opportunity ranking, competitive risks, and recommended actions. > Using the AgenticFlow CLI, deploy the market-intelligence-brief-pack blueprint. Research topic: "AI workflow automation demand in regulated industries" > > Market signals: "Southeast Asia: banks and media teams are piloting AI workflow tools but worry about compliance. US: operations teams ask for human-in-loop governance and audit trails. EU: buyers prioritize privacy-first deployments and model governance." > > Deploy, run once, show me the executive summary, regional insights, opportunity ranking, competitive risks, and recommended actions. > > Leave the workflow deployed. Print the Web UI link. *** #### 3. Stock Price Target Brief (`stock-price-target-brief`) Turn a single ticker and an investment style into a consistent research brief: quote snapshot, key news drivers, a watch rating, a risk checklist, and the next data to pull. Research and organization only — not financial advice. > Using the AgenticFlow CLI, deploy the stock-price-target-brief blueprint. Use these inputs: > > Ticker: AAPL Company name: Apple Investment style: conservative long-term watchlist > > Deploy, run once, show me the quote snapshot, key news drivers, watch rating, risk checklist, and next data to pull. > > Leave the workflow deployed. Print the Web UI link. *** #### 4. Sales Call Objection Coach (`sales-call-objection-coach-pack`) Turn a raw sales-call transcript into coaching: an objection table, the discovery questions the rep missed, clear next steps, and a Slack-style coaching feedback draft. Draft analysis only; it does not retrieve, send, or write to external systems. > Using the AgenticFlow CLI, deploy the sales-call-objection-coach-pack blueprint. Use this call transcript: > > "AE: Thanks for joining. What are you trying to improve? Prospect: Our support team spends too much time tagging tickets, but I am worried implementation will take months. AE: We can start with a small workflow. Prospect: Budget is tight this quarter and security will need to review any AI vendor. AE: I can send a case study. Prospect: Send it, but I need a clear rollout plan before involving my VP." > > Coaching policy: "Identify objections, buyer concerns, missed discovery questions, recommended next steps, and Slack-style coaching feedback. Draft analysis only; do not claim Fireflies retrieval, Google Drive file creation, Slack send, wait step, or external action occurred." > > Deploy, run once, show me the objection table, missed discovery questions, next steps, and Slack-style coaching feedback draft. > > Leave the workflow deployed. Print the Web UI link. *** #### 5. Maps Lead Enrichment (`maps-lead-enrichment-pack`) Turn a list of local businesses into a prioritized, enriched lead list — each business scored on the signals that matter, with a suggested outreach angle. Organizes the data you provide; no live scraping. > Using the AgenticFlow CLI, deploy the maps-lead-enrichment-pack blueprint. Use these local business rows: > > Each row: name, category, rating, number of reviews, website, city. > > Deploy, run once, show me the enriched and prioritized lead list — each business scored with the signals that matter and a suggested outreach angle. > > Leave the workflow deployed. Print the Web UI link. *** ### Get Started * **Try AgenticFlow free:** * **Latest changelog:** * **Join our Discord:** * **Found a bug?** Email with a reproducible case.