This week Anthropic released Claude Fable 5 — the most powerful model it has ever made public, now on everyone’s plan. The real story isn’t the benchmark sweep, it’s the safety valve wired inside it, quietly routing its own most dangerous questions down to a weaker model. Around it: open models small enough to live on your laptop, a fresh take on Claude Code out of Europe, and one almost-anonymous French engineer reminding us that a single person with a compiler can still leave a dent the size of the whole internet.

Topic of the Week

Claude Fable 5 goes public

What happened. On June 9 Anthropic put Claude Fable 5 in everyone’s hands — the most powerful model it has ever released publicly, and a “safe-for-general-use” version of its locked-down Mythos model. You feel the jump most on the long, messy, multi-step work that used to grind teams down. The story that made the rounds: Stripe handed it a code migration that normally eats a couple of months of engineering, and it was done in a day. It’s free to try on Pro, Max and Team plans until June 22, then it moves to usage credits.

What people are building with it. Within a day, timelines filled with one-prompt demos that are genuinely hard to believe: a playable Minecraft clone with biomes, ores and a day-night cycle in ~20 minutes, a working Swiss watch escapement in Three.js (real gear ratios, a breathing hairspring, hands showing the actual time), a cloned Windows desktop down to Solitaire and Edge, and a humanoid-robot design draft that ate ~1.4 million tokens in two hours. One thing to keep in mind: not all of it is real. A few of the most-shared clips turned out to be fakes — one person passed off old GTA-6 footage as Fable’s work — and The Register spotted the opposite problem too: Fable 5 sometimes refuses completely harmless prompts. Fun to scroll through, just don’t believe everything.

The twist worth noticing. Fable 5 ships with safeguards built into the model, not bolted on around it. On sensitive cybersecurity, biology and chemistry requests it doesn’t refuse — it silently falls back to Opus 4.8 to answer, and Anthropic says that fallback triggers in under 5% of sessions. External red-teamers spent 1,000+ hours hunting for a universal jailbreak and found none (the UK’s AI Safety Institute made partial progress). The unsafeguarded version — Claude Mythos 5, same underlying model with the guards lifted — is not on general release: it goes only to vetted cyberdefenders and a few biology researchers through Project Glasswing, a program run with the US government.

Why it matters. This is the cleanest example yet of a lab trying to ship frontier capability and frontier caution in the same product. A model that routes its own dangerous queries to a less capable sibling is a genuinely new design pattern — and it lands days after Anthropic itself warned (the Favaro/Clark post we closed on last week) that AI building AI could make humans lose control. The practical read for anyone evaluating models: Fable 5 is the new ceiling for hard, long-horizon work, but the safeguards mean its behaviour on edge-case prompts won’t always be the “real” Fable 5 answering. Worth knowing before you wire it into anything.

Fresh Papers

This week’s two papers rhyme: the hard part of building a good agent isn’t the model — it’s the environment you train it in.

DeNovoSWE — small models that build whole repos. Nobody has much training data for “here’s a spec, now write the entire codebase,” so this team built a pipeline to manufacture thousands of examples — each a real, working repository. Train a mid-sized open model on that and it goes from barely functional to near-frontier at building projects from scratch. The point worth keeping: a self-hostable model can get surprisingly close to the big names on greenfield work — no frontier API bills, no shipping your code to anyone.

Agentic Environment Engineering — a survey that names the discipline. It treats the sandbox an agent works in as a real engineering problem in its own right, and floats “Environment-as-a-Service” as the next step. Read next to last week’s harness-tree paper, the drumbeat is clear: the leverage is shifting from clever prompts to the scaffolding around the model.

New Models

Gemma 4 12B — the laptop-sized sequel. Remember the Gemma 4 31B deep-dive in #014, where the verdict on consumer hardware was “unusable” on a 16GB card? The 12B is Google’s fix for exactly that — an Apache-2.0 multimodal model (text, images, audio, video; 256K context) built to actually run on a 16GB laptop, reportedly near the bigger 26B at half the memory. If it holds, the “Claude plans, Gemma builds” loop we sketched in #014 could finally run on your own machine, not an H100.

Mistral Vibe — Europe’s answer to Claude Code. Mistral turned Le Chat into a coding agent: a plan-and-execute “Work mode,” sandboxed agents that open PRs, a CLI and a VS Code extension. Roughly Sonnet-level on coding benchmarks at about half the cost — and open enough to self-host, the data-sovereignty angle the US labs don’t offer.

Claude Code & Coding AI

The releases (v2.1.163 → v2.1.174). Fable 5 landed directly in Claude Code (v2.1.170). Two changes stand out for heavier users: nested subagents (v2.1.172) — subagents can now spawn their own subagents up to 5 levels deep — and a fallbackModel setting (v2.1.166) that lets you list up to three backup models Claude tries in order when the primary is overloaded, with an automatic one-shot retry. There’s also a new --safe-mode flag to launch with all customizations disabled, and /cd to move a session’s working directory without nuking the prompt cache. (Changelog.)

Worth a read. Anthropic published the explainer for dynamic workflows (the feature we flagged a couple weeks back). It walks through six reusable patterns — fan-out-and-synthesize, adversarial verification, tournament, loop-until-done and more — for when you want Claude to write its own orchestration harness instead of leaning on the default one.

Tools of the Week

Gemini 3.5 Flash Live Translate — speech-to-speech that doesn’t wait its turn. Google’s new real-time translation model covers 70+ languages and, instead of waiting for you to finish a sentence, generates translated speech continuously — staying just a few seconds behind and keeping your own intonation and pacing. It’s in public preview via the Gemini Live API and rolling into the Translate app; a Google Meet private preview can handle 2,000+ language combinations in a single meeting. The use case Google points to: Grab’s 10M+ monthly driver-rider voice calls.

A tidier Bedrock console. Amazon Bedrock shipped a new console optimized for Anthropic- and OpenAI-compatible APIs, making model selection and deployment less fiddly — small, but it’s the plumbing that decides how easily a regulated client can actually adopt these models.

AI at Tenvalleys

We say it internally often enough that it’s worth putting in writing: we think the real backbone of AI transformation isn’t the enterprise rollout — it’s education. The people who’ll build with these tools for the next thirty years should grow up AI-fluent, not get retrofitted at 30. That belief is why our fully pro-bono engagement goes to Wiśniowa Technikum in Warsaw, where we’re helping the faculty rebuild the “Technik Programista” track into an AI-native “Programista AI” curriculum, launching September 2026.

This Monday the school hosted a conference, and our CEO Daniel Bachan represented us on stage with a talk on how a programmer’s job has changed in the age of AI. The through-line: the programmer stopped being a machinist typing code line by line and became an architect, conductor and navigator — what stopped mattering is the typing; what became priceless is knowing what to build, how to steer the machine, and how to navigate complexity no single person can hold in their head.

Education like this works best as a team sport. If you’re building in this space — a school, a company, anyone who wants to help shape what AI-native vocational training looks like — we’d genuinely love to collaborate. It’s bigger than any one company, and the more hands on it, the better. Reach out.

For Dessert

A post about Fabrice Bellard went viral again this week, and it’s worth the detour. He’s a French programmer who keeps almost no public presence — no social media, no interviews — while his code quietly runs a huge slice of the internet. He wrote FFmpeg, the engine that decodes and encodes video behind basically every YouTube clip, Netflix stream and VLC window. Then he wrote QEMU, the emulator that underpins enormous amounts of cloud virtualization. Then, as if those weren’t each a career: a tiny self-hosting C compiler that can compile and boot the Linux kernel in seconds, the QuickJS JavaScript engine, a couple of image and video codecs — and for fun, in 2009, he computed pi to 2.7 trillion digits and broke the world record on a single home PC that cost under $3,000. In a week where the headline is a model that can rewrite 50 million lines of code in a day, it’s a nice reminder of what one stubborn human with a compiler can still pull off.

Prepared at Tenvalleys — a delivery-first AI engineering partner — by Nikola Powałka. Feedback? Email us at contact@tenvalleys.com or reach out on LinkedIn.