🤖 Claude Fable 5 Dropped. Here's the Split.

Anthropic shipped two models today under different names and different terms. Claude Fable 5 is the public release: a Mythos-class model, Anthropic says, capable of long-running multi-step work end to end. Claude Mythos 5 goes to trusted partners only, specifically cyber-cleared organizations, and comes without the safety constraints Anthropic applied to the public version.

That split is the story. Fable 5 is what you get. Mythos 5 is what exists.

Vercel already has Fable 5 on AI Gateway, available to route to right now without any model-access waitlist. If you're running coding agents or long-context pipelines, today is the day to benchmark it against your current setup.

The Drops below have several repos that pair well with it.

---

[Skill] claude-ads, A comprehensive paid advertising audit and optimization skill for Claude Code. Runs 250+ checks across Google, Meta, YouTube, LinkedIn, TikTok, Microsoft, and Apple Ads with weighted scoring. If you run paid acquisition, this is the audit you would pay an agency for, now running inside your terminal.

[Skill] claude-code-prompt-improver, A pre-send hook that intercepts your vibe-typed instructions and sharpens them before Claude Code executes. 1,599 stars. The gotcha it prevents: low-specificity prompts producing six rounds of back-and-forth before Claude lands on what you meant.

[Repo] guizang-social-card-skill, A Claude Code skill that generates Xiaohongshu carousels and square cover pairs from a single prompt. 3,221 stars. 28 layouts, 10 themes, single-file HTML to PNG output. If you produce social content in quantity, this compresses one hour of design into a command.

[Repo] claude-code-auto-memory, A Claude Code plugin that automatically maintains your CLAUDE.md files as you work. 148 stars. The problem it solves: project memory going stale because you never updated the file. It watches, writes, and keeps the context current.

[Repo] roboflow/supervision, A reusable computer vision toolkit: detection, tracking, annotation, and zone logic in one library. 42,929 stars and GitHub trending now. If you're wiring vision models into an agent pipeline, this is the layer between the model output and whatever you want to do with it.

[Repo] aiwaves-cn/agents, An open-source framework for data-centric, self-evolving autonomous language agents. 5,930 stars. The self-evolution hook is the interesting part: agents that update their own behavior from feedback rather than requiring manual prompt revision.

[Repo] Yifan-Song793/RestGPT, An LLM-based autonomous agent that controls real-world applications via RESTful APIs. 1,399 stars. If you have a SaaS tool with an API and no native agent integration, this is the fastest path to making it agentic.

[Repo] nari-labs/dia, A TTS model that generates ultra-realistic dialogue in a single pass. 19,308 stars. One-pass generation means no stitching, no latency stacking across segments. Relevant if you're building voice interfaces or audio content pipelines. It wants a real GPU: if your box can't carry it, rent one on Runpod by the hour.

[Repo] google-gemini/gemini-cli, An open-source AI agent that runs Gemini directly from your terminal. 105,097 stars. Worth having alongside Claude Code as a secondary inference path, especially for tasks where Gemini's context window or pricing makes more sense.

---

[MCP] xuanlinAI/overmind, A 66-module, 6-channel AI cognitive system designed as a single persistent brain across all your AI tools. The specific operator value: cross-session memory and multi-agent communication that survives context resets. Zero-config claim. If you've been manually re-briefing Claude on project state at the start of every session, this is the architectural fix.

---

Claude Fable 5 is live on Vercel AI Gateway. Anthropic's new public model is available now via Vercel's AI Gateway, no access request needed. Fable 5 is described as Mythos-class on long-running, ambiguous, multi-step tasks. Operators running coding or reasoning agents should benchmark it today.

Anthropic split the new Claude release by trust tier. Wired reports Mythos 5 goes only to cyber-cleared organizations; Fable 5 is the public version. The operator consequence: the most capable variant of Claude is now gated behind a partner vetting process you cannot self-enroll in. Plan your capability roadmap accordingly.

Apple's Xcode 27 ships agentic coding. Apple announced new AI frameworks and extended agentic coding in Xcode 27, including new APIs for integrating AI models directly into the build loop. If you ship iOS or macOS apps, your Xcode workflow is about to change whether you opt in or not.

Simon Willison on AgentsView for token cost tracking. Willison wrote up AgentsView, a tool by Wes McKinney for exploring token usage across multiple coding agents running locally, including support for custom model pricing. Useful if you're running mixed-model setups and want actual cost visibility rather than estimating.

FrontierCode ships a code-quality benchmark. Latent Space released FrontierCode, a benchmark targeting code quality over output volume, specifically to distinguish signal from slop in coding agent evaluations. Worth watching: if it gains adoption, it becomes the standard your agents will be measured against.

---

This week's technique: Subagents and orchestration. One parent agent, specialized children, and the routing logic that makes them cheaper to run.

Most operators run Claude Code as a single agent handling everything in one context window. That works until the task is large enough that the window fills, quality drops, and you're paying for tokens that aren't doing anything. Subagents fix this.

The pattern: a parent agent receives the task, decomposes it, and dispatches specialized subagents to handle scoped subtasks. Each child runs in a fresh, focused context. You get cleaner outputs, lower per-task cost (route cheap subtasks to a smaller model), and the ability to run subagents in parallel.

1. In your parent agent session, use the Task tool to spawn a subagent: give it a specific role, a scoped instruction, and only the context it needs. Nothing extra. 2. For parallel work, dispatch multiple Task calls before awaiting any of them. Claude Code runs them concurrently by default. 3. Route cost-sensitive or repetitive subtasks (summarization, formatting, simple classification) to a lighter model by specifying it in the subagent config.

You'll know it's working when your parent context stays tight and each child returns a clean, scoped output rather than a bloated scratchpad.

Claude Code docs: sub-agents

---

Move: Auto-audit a Git diff before you push.

Connect Claude Code to your repo with the GitHub MCP, then add a pre-push hook that pipes the staged diff to Claude with a single prompt: "Review this diff for logic errors, security issues, and missing edge cases. Flag only what matters."

1. Install the GitHub MCP and confirm it's connected with /mcp in-session. 2. Add a .claude/hooks entry: trigger on pre-push, pass the output of git diff HEAD~1 as the prompt input. 3. Set the response format to a numbered list. Anything Claude flags, you review. Anything clean, it ships.

You'll know it worked when Claude catches the off-by-one or the unchecked null before the CI pipeline does, and you stop apologizing in Slack for the same class of bug twice.

---

This Friday's kit is Ad Factory: how to produce a talking-head founder ad using an AI avatar, no camera, no studio, and nothing that reads as generated. Most solo operators skip video ads entirely because production is the bottleneck. The kit removes that constraint. Grab operator access before Friday.

---

Fable 5 is live, Mythos 5 is gated, and the gap between them is already a strategic variable. That's the move to watch this week.

This issue contains affiliate links; we may earn a commission.

See you Thursday.