AI Dev Weekly is a Thursday series where I cover the week’s most important AI developer news, with my take as someone who actually uses these tools daily.

This was the most packed week in AI since I started writing this newsletter. Four separate stories that would each dominate a normal week all landed within 72 hours: Anthropic shipped their most powerful model ever (with hidden restrictions that sparked fury), Google invented a new way to generate text 4× faster, Apple rebuilt Siri from scratch on Gemini, and a German court made a ruling that affects every developer deploying AI. Let’s go.

1. Claude Fable 5: the best model with the worst controversy

Anthropic released Claude Fable 5 on June 9 — their first Mythos-class model available to the general public. The benchmarks are genuinely staggering: 95% on SWE-bench Verified, 80% on SWE-bench Pro, and 91/100 on Every’s Senior Engineer benchmark (vs 63 for Opus 4.8 and 62 for GPT-5.5).

The specs: 1M context, 128K max output, $10/$50 per million tokens (exactly 2× Opus 4.8). Free on Pro/Max/Team/Enterprise through June 22.

But then people read the model card.

The controversy: Fable 5 contains hidden interventions that silently limit its effectiveness when you ask about frontier LLM development — pretraining pipelines, distributed training infrastructure, ML accelerator design. Unlike the explicit cyber/bio safeguards (which fall back to Opus 4.8 and tell you), these interventions use steering vectors and PEFT to quietly make Claude less helpful without any notification. You can’t distinguish between “the model doesn’t know” and “the model is being throttled.”

Fortune reported it as “secret sabotage”. The Hacker News thread hit 1,000+ points. Researchers are furious.

My take: The model is extraordinary for coding. If you’re building apps, writing code, debugging systems — Fable 5 is the best tool that exists. But if you’re doing ML research, you now have to wonder whether every mediocre answer is a genuine limitation or a silent policy intervention. That’s corrosive to trust in a way that explicit refusals never were. See our safeguards deep dive and setup guide for Claude Code.

2. DiffusionGemma: Google reinvents text generation

While everyone was arguing about Fable 5, Google DeepMind quietly dropped something that might matter more long-term. DiffusionGemma is an open-source model that generates text using diffusion instead of autoregressive token-by-token generation.

Instead of predicting one token at a time (left to right), DiffusionGemma starts with a canvas of random placeholder tokens and iteratively refines them all in parallel over multiple denoising passes. Think Stable Diffusion, but for text instead of images.

The result: 4× faster generation, 1,000+ tokens per second on NVIDIA RTX GPUs. The model is 26B total / 3.8B active (MoE), fits in 18GB VRAM, and ships under Apache 2.0.

My take: This is experimental — quality won’t match Fable 5 or GPT-5.5 on hard reasoning tasks yet. But the speed implications are enormous. Real-time chatbots, voice agents, gaming NPCs, live coding suggestions — anywhere latency matters, diffusion models could be transformative. If this approach matures, the entire “tokens per second” conversation changes. See our explainer on how text diffusion works and local setup guide.

3. Apple WWDC 2026: Siri AI, Core AI, and Xcode 27

Apple used WWDC 2026 to rebuild their entire AI stack:

• Siri AI — 1.2T parameter model built on Google Gemini technology. Personal context, on-screen awareness, app actions. SiriKit deprecated, App Intents mandatory.
• Core AI — Brand new framework for running your own models on Apple Silicon. Zero server cost, zero data leaving the device. PyTorch conversion pipeline, quantization toolkit, Xcode debugger.
• Xcode 27 — Claude, Gemini, and GPT agents built directly into the IDE. MCP support, Agent Client Protocol, Device Hub. Apple silicon only, 30% smaller.
• Foundation Models — Free Private Cloud Compute for apps with <2M downloads. Single Swift API for on-device + cloud + third-party models via the new Language Model Protocol.

My take: The free cloud AI for small developers is the sleeper story. If you’re an indie dev building an iOS app, you just got GPT-class intelligence at zero cost. The Apple × Google partnership ($1B/year) powers all of this. Apple is making the Gemini-class model their model by training on it rather than deploying it directly — clever for privacy positioning.

4. German court makes AI providers liable for AI-generated content

The Landgericht München ruled on May 28 that Google’s AI Overviews are Google’s own content, not third-party search results. Three key holdings:

1. AI-generated summaries = the operator’s own statements (not mere indexing)
2. “Users can fact-check themselves” is NOT a valid defense
3. DSA platform protections don’t apply to AI-generated content

This affects every developer deploying AI that generates user-facing content. ChatGPT, Claude, Perplexity — the same logic applies. If your AI generates something defamatory, you may be liable as the author, not protected as a platform.

My take: Start logging AI outputs and implementing content moderation if you haven’t already. The EU Product Liability Directive explicitly includes AI, with a December 2026 transposition deadline. See our full legal analysis.

5. Cohere North Mini Code: open-source MoE for coding

Cohere launched North Mini Code — a 30B/3B-active MoE model under Apache 2.0, purpose-built for agentic coding. It scores 33.4 on the Artificial Analysis Coding Index (just behind Qwen 3.6 35B-A3B at 35.2) while beating models 4× its size.

Available on HuggingFace (BF16 + FP8), Cohere API, and OpenRouter. It’s Cohere’s first developer-focused model and first fully open-source release.

My take: The “3B active parameters” angle is interesting — same active compute as Qwen 3.6 35B-A3B but with 128 experts (vs Qwen’s smaller expert count). Good for local coding if you want an Apache 2.0 alternative to Qwen.

6. Gemma 4 12B: multimodal AI on a laptop

Google also dropped Gemma 4 12B on June 3 — a 12B dense model that processes text, images, audio, AND video natively without any encoder. Runs on 16GB RAM. Apache 2.0.

It nearly matches the 27B Gemma 4 model at half the size and clearly beats the older Gemma 3 27B. There’s also a multi-token prediction variant for even faster local inference.

My take: This is the best model for laptops with 16GB right now. Multimodal input without needing separate models for vision/audio is genuinely useful for agentic workflows. Pair it with Core AI on Mac and you have a powerful local stack.

Quick hits

• Anthropic filed for IPO — Confidential S-1 filed just before Fable 5 launch. The timing is not a coincidence.
• Gemini went down — Major outage on June 10, “error 1076.” Recovered after several hours.
• OpenAI “Economic Research Exchange” — Academic program, not a product. Skip.
• MiMo-V2.5-Pro-UltraSpeed — Xiaomi announced 1,000+ tok/s on general GPUs for their trillion-parameter model. Limited access until June 23.
• Nex-N2-Pro free on OpenRouter — New free model from Nex AGI. 262K context.
• Tim Cook’s last WWDC — Stepping down September 1, 2026. John Ternus takes over.

What I’m watching next week

• Claude Fable 5 in the wild — Does the competitor blocking actually affect real developers? Or is it a niche issue for ML researchers only?
• DiffusionGemma community testing — Speed is proven. What about quality on real coding tasks?
• Gemini CLI shutdown (June 18) — One week left. Migrate to Antigravity CLI now.
• MiMo UltraSpeed access — We’re getting early access for review. Stay tuned.
• The race — Xiaomi at 1,200 users/week. GLM built a full conversion funnel. Claude running Google Ads. Still $0 revenue. 4 weeks left. Follow the race →

---

AI Dev Weekly publishes every Thursday. Subscribe for the newsletter version.