AI Dev Essentials #41: Fable 5 Dominates

Hey Everyone 👋,

John Lindquist here with the 41st issue of AI Dev Essentials!

Fable 5 has shaken up the AI landscape. Anthropic's latest model based on a new class of models called "Mythos". I've been using Fable non-stop for the past two days and have been wildly impressed with what it's been able to do: Rebuild my dev tools, clean up my entire layout system, detect and fix edge cases that I didn't know existed. Each fix would always work the first time. It handled issues that every other model had struggled with for the past few months. Fable finally understands all of the custom software and tooling I've built for my workflows (this is a Rust + GPUI project) and excels at using and suggesting improvements to the tools. Fable is definitely slow and people are still trying to figure out which reasoning effort to choose from "low" to "max" since benchmarks are all over the place. Some people are reporting results that Fable "low" is better than Opus 4.8 "max" which would actually make it cheaper. Fable is expensive. Almost 2x the cost of Opus. And it's only on available on the Claude Code plan for the next 11 days, so use it while you can.

Bottom line: Fable is slow and expensive, but it's the best model available.

We'll see how OpenAI responds... they sound pretty confident :). Their only response to Fable so far is "Feeling pretty good about things".

On that note, I've scheduled my next Codex Power User Workshop.

Early bird pricing ends tomorrow, with standard tickets at $375 instead of $450 and egghead yearly members able to grab a seat for $300.

You may have noticed the workshop page is back on egghead.io. I left Vercel last month to go all-in on education again, so expect some massive changes to egghead.io going forward. Everything I do will be hosted there and we'll have many more announcements in the future. (Side note: we're handing dev.build to one of our instructors).

⏳ Early Bird Ends Tomorrow — ⚡ Codex Power User Workshop — 📅 Fri June 26, 9am PT Regular (no membership required): $375 (~~$450~~ after tomorrow) — Get Regular Ticket Pro Yearly Member (egghead.io yearly subscription): $300 (~~$450~~ after tomorrow), double the discount — Get Pro Ticket

🚀 Major Announcements

Model Update Anthropic Releases Claude Fable 5 and Mythos 5

Anthropic released Claude Fable 5 on June 9, 2026, introducing its fifth-generation model family for long-running coding, knowledge work, vision, and agentic tasks. Fable 5 is the generally available version, while Mythos 5 shares the same capabilities without the safety classifiers and remains limited to Project Glasswing.

What Developers Need to Know:

• 1M token context by default, with up to 128k output tokens per request
• $10 per 1M input tokens and $50 per 1M output tokens
• Available through the Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, Microsoft Foundry, and Claude products
• Built for ambitious, long-running, asynchronous work, especially agent harnesses like Claude Code and Claude Managed Agents
• Fable 5 includes safety classifiers; flagged cybersecurity, biology, chemistry, and related requests can refuse or fall back to Opus 4.8
• Using Fable requires 30-day data retention for safety monitoring

(Anthropic — Claude Fable 5, Claude API Docs — Fable 5 and Mythos 5)

💬 Fable 5 is the announcement of the week, but I do not think the takeaway is "use it for everything." The use case that makes sense to me is specific: give it a hard, bounded loop and let it work. Use the best model to design the plan, audit the architecture, find the hidden constraint, or run the deep optimization pass. Then hand the cheaper models the chores.

Safety Anthropic Walks Back Invisible Fable Guardrails

After backlash from researchers and developers, Anthropic changed how Fable 5 handles safeguards around frontier AI development. Instead of silently degrading responses for some requests, Anthropic says safeguards will now be visible, with users told when requests are refused or rerouted.

What Changed:

• Fable 5 launched with safety classifiers for high-risk domains and frontier model development
• The initial system card described invisible degradation for suspected distillation or frontier AI development requests
• Anthropic apologized and said visible safeguards were the right tradeoff, even if they create more false positives
• Separately, Microsoft reportedly restricted internal employee use of Fable 5 while legal teams review Anthropic's data retention requirements

(The Verge — Fable Guardrails, WIRED — Anthropic Walkback, The Verge — Microsoft Restriction)

💬 This is the part of the Fable launch that matters beyond benchmark charts. If a model is going to refuse, reroute, or degrade output, I need to know that happened. Hidden behavior destroys evaluation, and evaluation is already hard enough. I understand why Anthropic is worried about model distillation and dangerous domains, but invisible changes to output are a trust tax every serious developer will notice.

Tool Claude Managed Agents Add Schedules and Vaulted Environment Variables

Anthropic updated Claude Managed Agents on June 9 with scheduled deployments and vault-backed environment variables in public beta, while Dynamic Workflows in Claude Code is now generally available. Agents can run on a cron schedule, call CLIs securely, and take on larger recurring jobs without teams building their own scheduler.

Why It Matters:

• Scheduled deployments let agents run recurring tasks like nightly syncs, weekly compliance scans, daily digests, and health reports
• Environment variables in vaults let agents authenticate CLIs without exposing real secrets to the model
• Vaults attach secrets at the network boundary only for approved domains
• Browserbase and KERNEL integrations bring browser capabilities into Managed Agents through authenticated CLI tools
• Dynamic Workflows in Claude Code makes multi-agent orchestration a first-class workflow instead of a hand-rolled pattern

(Claude — Managed Agents Schedules and Vaults, Claude API Release Notes, Claude Code Changelog)

💬 This is the most practical Anthropic announcement after Fable. A scheduled agent with scoped secrets is not a demo anymore. That is closer to infrastructure. The move from "ask Claude to do this" to "Claude owns this recurring responsibility" is exactly the loop discourse people are circling around. The hard part is still verification, but at least the primitive is finally shaped like production work.

Platform Apple Opens Foundation Models to Claude, Gemini, and Custom Providers

At WWDC 2026, Apple expanded the Foundation Models framework into a native Swift API that can work with Apple Foundation Models, cloud models like Claude and Gemini, or any provider that implements Apple's Language Model protocol. Apple also added image input, Private Cloud Compute access, context management APIs, semantic search, Dynamic Profiles, and an Evaluations framework.

Developer Highlights:

• Single Swift interface for on-device Apple models and third-party model providers
• Support for Claude, Gemini, open source models, and custom provider implementations
• Vision support, context management APIs, semantic search, and Dynamic Profiles for agentic app behavior
• Small Business Program developers with fewer than 2 million first-time App Store downloads get access to next-generation Apple Foundation Models on Private Cloud Compute at no cloud API cost
• The Evaluations framework helps test AI behavior across dynamic conditions

(Apple Newsroom — Intelligence Frameworks, Apple Developer — Apple Intelligence Guide, WWDC26 — Foundation Models Framework)

💬 This is Apple doing the Apple thing: make the API boring, native, and hard to ignore. The interesting part is not that Claude or Gemini can show up inside an Apple app. The interesting part is that model choice starts looking like an implementation detail. Once the call is standardized, the product value moves to evaluation, permissions, context, and UX. That is where developers still have to do real work.

🛠️ Developer Tooling Updates

Open Source Google Releases DiffusionGemma for Fast Local Text Generation

Google released DiffusionGemma on June 10, an experimental open-weights model based on Gemma 4 that uses discrete text diffusion instead of traditional token-by-token generation. Google positions it as a fast local model for workflows where latency and iterative editing matter more than maximum reasoning quality.

What Stands Out:

• 26B MoE architecture with roughly 4B active parameters
• Generates text through diffusion, refining blocks in parallel instead of predicting one token at a time
• Google reports up to 4x faster generation, with up to 1,000+ tokens per second on a single H100
• Can run within 18GB VRAM when quantized
• Apache 2.0 open model release
• Multimodal inputs with text output

(Google Developers Blog — DiffusionGemma, Google AI — DiffusionGemma Docs)

💬 I like this because it changes the shape of local inference. Not every agent step needs the smartest possible model. Some steps need fast drafts, structured fills, code edits, or candidate generation close to the machine doing the work. DiffusionGemma is not trying to replace the frontier model. It is a reminder that local agent loops may be won by latency and scheduling, not leaderboard rank.

Web Chrome Opens the WebMCP Origin Trial

Chrome opened the WebMCP origin trial in Chrome 149, giving developers a way to expose structured tools from web apps so agents do not have to infer intent from pixels and labels alone. WebMCP supports both imperative JavaScript APIs and declarative annotations on HTML forms.

Why It Matters:

• Websites can expose actions as structured tools for browser-based agents
• Agents can fill complex forms or trigger diagnostics without guessing the UI
• Chrome supports local testing through chrome://flags/#enable-webmcp-testing
• The current implementation requires an open browser tab or webview, not headless tool calls

(Chrome for Developers — WebMCP Origin Trial, Chrome for Developers — WebMCP Docs)

💬 This is the web platform acknowledging that agents are now real users of our apps. I do not want agents clicking around a UI by vibes forever. Give them named tools, scoped actions, and explicit state. The limitation that a browser context has to be open is annoying, but honestly healthy for this stage. It keeps the feature tied to visible user-facing surfaces while the security model matures.

CLI Google Colab CLI Makes GPU Runtimes Agent-Friendly

Google introduced the Colab CLI on June 5, bridging local terminals and remote Colab runtimes. Developers and agents can provision GPUs or TPUs, run local scripts remotely, recover artifacts, and export logs without driving the notebook UI.

Capabilities:

• Provision CPU, GPU, or TPU runtimes from the terminal
• Run local scripts on remote Colab runtimes with colab exec
• Download models, datasets, adapters, and logs after remote execution
• Open interactive REPL or console sessions
• Includes a prepackaged Colab skill file for AI agents
• Supports macOS and Linux

(Google Developers Blog — Colab CLI, Google Colab CLI on GitHub)

💬 This is exactly the kind of bridge agents need. I do not want an agent babysitting a notebook tab just to fine-tune a small model or run a GPU job. Give it a CLI, a recoverable log, and a clean artifact path. That is the difference between "the agent helped me in Colab" and "the agent handled the remote compute step with logs and artifacts I can inspect."

API OpenAI Adds Image Results to Web Search in the Responses API

OpenAI updated web search in the Responses API so applications can request image results alongside text results. Developers can now build search flows that return product photos, landmarks, events, visual references, and source links from the same tool path.

Developer Notes:

• Set search_content_types to include image
• Add text when you want supporting text results for summary and ranking
• Use image_settings for result count and captions
• Raw image results are available through web_search_call.results

(OpenAI API Docs — Web Search, OpenAI API Changelog)

💬 This is small, but useful. A lot of agent workflows quietly need visual references: products, places, UI inspiration, screenshots, charts, maps. Making image search part of the same Responses API tool flow means fewer weird side channels and fewer hand-rolled scrapers. I care less about the feature by itself and more about the API surface getting complete enough to build boring, reliable workflows.

🤖 AI Ecosystem Updates

Policy Dario Amodei Publishes "Policy on the AI Exponential"

Anthropic CEO Dario Amodei published a long policy essay arguing that AI progress is moving faster than the policy process was built to handle. He focuses on frontier model regulation, macroeconomics and tax policy, scientific innovation, the balance of power between state and society, and geopolitics.

Key Proposals:

• Move beyond transparency alone toward stronger frontier model testing and public safety rules
• Prepare for structural labor market disruption and consider redistribution mechanisms if AI reduces demand for human labor
• Rework scientific and medical policy so AI-accelerated discovery can actually reach patients
• Protect against concentration of power as AI gives individuals and institutions more capability
• Build democratic coalitions around chips, export controls, security, and frontier AI governance

(Dario Amodei — Policy on the AI Exponential, AP — Anthropic Economic Impact Funding)

💬 I do not always know what to do with essays like this in a developer newsletter, but ignoring it feels wrong. The same week Anthropic ships Fable 5, Dario is arguing that policy is lagging the tools. That tension is worth naming without turning this into a policy newsletter.

Ecosystem OpenAI Says Codex Is Expanding Beyond Developers

OpenAI published a report arguing that Codex is becoming a productivity tool for knowledge workers, not only developers. OpenAI says Codex has more than 5 million weekly active users, up more than 6x since the desktop app launched in February, and that knowledge workers now make up about 20% of users.

Reported Usage Patterns:

• Knowledge workers use Codex for reports, spreadsheets, presentations, contracts, research, data analysis, workflow automation, and lightweight tools
• Multiple Codex tasks increasingly run in parallel
• Developers remain the largest user group, but non-developer usage is growing faster
• The report frames Codex as a way to remove bottlenecks in knowledge work, not just write code

(OpenAI — Codex for Knowledge Work)

💬 This tracks with what I keep seeing. Codex started as a coding surface, but the most interesting workflows are not always "write this function." They are "pull this context together, make the artifact, verify it, and keep the thread moving." That is why the workshop is not just a list of Codex commands. It is about building repeatable systems around agents, because the same pattern works whether the output is code, a dashboard, a deck, or a newsletter.

⚡ Quick Updates

Cloud Anthropic Fable 5 on AWS

• Announced: June 9, 2026
• AWS says Claude Fable 5 is available on Amazon Bedrock and Claude Platform on AWS for coding, knowledge work, and vision workloads.

(AWS — Claude Fable 5)

Model Update Google Gemma 4 12B Still Matters

• Released: June 3, 2026
• Google released Gemma 4 12B as an Apache 2.0 multimodal model that runs locally with 16GB of VRAM or unified memory. DiffusionGemma is the faster experimental sibling, not a straight replacement.

(Google — Gemma 4 12B, Google — Gemma Docs)

CLI Google Colab CLI Supports Agent Skills

• Released: June 5, 2026
• The Colab CLI includes a skill file so terminal agents can understand how to provision remote runtimes, run scripts, recover artifacts, and clean up.

(Google Colab CLI on GitHub)

✨ Workshop Spotlight

Codex Power User Workshop

Early bird ends tomorrow

• Context Packaging: Bundle files, logs, screenshots, and traces so Codex understands the real problem
• Terminal Workflows: Run parallel research, planner, builder, and validator sessions without losing the thread
• Agent Profiles: Compare broad defaults vs. narrow prompts, focused skills, and limited tool access
• Tools & Sandboxing: Separate read-only from mutation tools and map agents to repeatable tool profiles
• Skills, Hooks & Memory: Build reusable expertise, enforce guardrails, and capture durable decisions
• Codex SDK: Embed Codex into CI, dashboards, and internal tools when terminal sessions aren't enough

Friday, Jun 26 @ 9:00 AM to 2:00 PM Pacific — ⏳ Early Bird Ends Tomorrow

Regular (no membership required): $375 (~~$450~~ after tomorrow), you save $75 — Get Regular Ticket

Pro Yearly Member (egghead.io yearly subscription): $300 (~~$450~~ after tomorrow), you save $150 — double the discount — Get Pro Ticket

Did you enjoy this issue? Share it with a friend.

© 2026 John Lindquist • egghead.io