Welcome back. For two months, Anthropic said their Mythos model was too dangerous to release. Now, it's finally here. Claude Fable 5 just dropped with Mythos-level power and new safety guardrails. But there is a catch. Those same guardrails are quietly deciding which model actually answers you behind the scenes. We dive into exactly how that works in today’s insight.

Also: How a Googler writes code for agents, the 3 biggest changes the Claude Code team saw working with Fable 5, and an engineer's guide to loop-based coding workflows.

Google rolls out live voice translation in 70 languages: The search giant just dropped Gemini 3.5 Live Translate. It converts speech to speech across 70+ languages while keeping the speaker's pitch, pacing, and intonation intact. The model auto-detects languages and holds up in noisy settings, staying just seconds behind the speaker. It's live worldwide in the Google Translate app on Android and iOS, with developer access in API preview.

Cohere drops its first open-source coding model: The Canadian AI lab just open-sourced North Mini Code, a mixture-of-experts design model built for agentic software work. It has 30 billion total parameters but only uses 3 billion at a time. This allows teams to run it on a single H100 GPU, meaning they can keep their data in-house and avoid being tied to a specific provider. Cohere says it hits up to 2.8x the output throughput of Devstral Small 2. Download the weights.

Apple unveils cross-company subscription bundles for the App Store: At WWDC, the iPhone maker laid out App Store changes that developers will want to track. Later this year teams will be able to sell subscription bundles and suites across different companies, which includes packaged offers that aren't available on their own. Apple is also dropping the Intel chip requirement for Mac apps and purging "opportunistic" listings.

Chatbots are just the beginning. Google for Startups’s Startup School: Agentic AI kicks off today, teaching you how to design and deploy autonomous systems that actually execute complex tasks. 

Master advanced tool calling, multi-agent orchestration, and complex reasoning chains directly from Google experts. Transform your AI features into scalable, independent workflows. 

Classes start today—save your spot now!

One launch hides two models. Anthropic shipped Claude Fable 5 yesterday. It is their first public "Mythos-class" model. Fable 5 uses the same weights as the restricted Mythos 5 model. The main difference is the wrapper. Fable runs safety classifiers on every request. If a classifier triggers, the request is handed to the older Opus 4.8 model instead — and on Claude's own surfaces, you're told when it happens. Because of this, Fable acts more like a routing system than a single model. It essentially decides which model should answer you based on safety rules.

The benchmark trap. The launch table groups both models into one column and only shows the highest scores (see the above image). Anthropic claims switching between models is rare, happening in less than 5% of cases. But when benchmarks trigger safety filters, these handoffs happen much more often. For example, on Terminal-Bench, nearly 21% of Fable trials were refused and then finished on Opus 4.8. This means the benchmark results show high-end performance, but your actual session might be running on a different model entirely.

Some routing you never see. Requests involving frontier-LLM development don't reroute to other models. Instead, Fable simply lowers the quality of its own output without telling you. Simon Willison, the co-creator of Django, noted that this is a first for Anthropic. Just one day after launch, a developer said Fable flagged his training code. It then quietly swapped in a completely different architecture.

The routing layer is your job now. Anthropic's own prompting guide confirms it:

It used to be about clearly explaining your task. Now, the challenge is describing it without sounding interesting enough to trigger a safety filter.

Codex's performance tends to drop during long sessions. This happens because every file it explores stays in its context, which eventually crowds out your actual task. OpenAI's documentation uses a read-only agent as a fix. This agent explores in its own thread and reports back a summary.

To set this up, create “.codex/agents/code-mapper.toml” in your project:

Codex only creates subagents if you specifically name them. Just tell it, "Have code_mapper trace user sessions and summarize the results." This scan runs on a cheaper model in a read-only sandbox, so your main thread stays clean and the subagent won't change any of your code.

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

How to run agent loops without wasting tokens: If you want to implement recursive loops without breaking the bank, this tutorial is a great place to start. You’ll learn the mechanics of agentic loops, why fully autonomous systems can drain your token budget, and how to use constrained loops to automate code reviews.

Mainframe: It turns any GitHub PR into a clear video walkthrough. It includes UI previews and smart context and helps you understand changes in seconds.

Career-Ops (52.2K ⭐): A job search tool for your AI coding agent. Just provide a job URL to get a custom PDF resume and a match score based on your CV. It logs everything locally in one command, keeping your data private.

Fine-tuning LLMs with the xLAM Dataset (by Hugging Face): LLMs often struggle to interact with external APIs or use structured function calls accurately. This cookbook builds a universal, automated pipeline using QLoRA and the xLAM dataset to efficiently fine-tune various open-source models on consumer-grade hardware.

This skill captures a former Vercel engineer's decade of teaching experience to build you a personalized learning path for any topic.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team