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 week felt like watching dominoes fall. On Friday the US government yanked Fable 5 and Mythos 5 from every non-American developer on the planet. By Tuesday, a Chinese lab had open-sourced a model that rivals them. Today, Gemini CLI officially dies. And GPT-5.6 is confirmed for Monday. Buckle up.

1. US government bans Fable 5 and Mythos 5

On June 12, the US Commerce Department issued an export control directive ordering Anthropic to immediately suspend all access to Claude Fable 5 and Mythos 5 for any foreign national, whether inside or outside the United States. That includes foreign Anthropic employees.

Within hours, Anthropic disabled both models for everyone worldwide to comply. Pro and Max subscribers were rolled back to Opus 4.8.

What happened: Anthropic’s own model card disclosed that Fable 5 scores 95% on SWE-bench Verified. Combined with the hidden ML-research throttling (which essentially means the model could help build competing frontier AI if unthrottled), the government apparently decided it’s too capable to let foreign nationals access. The irony: Anthropic’s transparency about capabilities may have triggered the ban that wouldn’t have happened if they’d been vague about benchmarks.

The fallout:

• Every developer outside the US lost access to the best coding model overnight
• Anthropic says they disagree with the order but complied immediately
• The G7 summit this week (see below) is partly a response to this chaos
• Anthropic’s IPO timeline just got complicated

My take: This is a watershed moment. If you’re a developer in Europe, Asia, or anywhere outside the US, you just learned that API access to frontier models can vanish with zero notice. The practical implication: you cannot build a business that depends on a single closed-source frontier model. You need fallbacks. Open weights just went from “nice to have” to “business continuity requirement.”

We covered this in detail: Fable 5 ban explained and what developers should do now.

2. GLM-5.2 drops open weights on the same week

The timing could not be more dramatic. On June 13, Z.ai (formerly Zhipu AI) launched GLM-5.2 on their coding plan. On June 17 (yesterday), they released full MIT-licensed open weights on Hugging Face.

The numbers: 753B MoE model, ~40B active parameters per query. Ranks #2 on Code Arena behind only Fable 5. Beats GPT-5.5 on long-horizon coding benchmarks. Within 1% of Opus 4.8 on agentic coding. 1M usable context. MIT license means you can self-host, fine-tune, and use commercially with zero restrictions.

The cost difference: $5.80 per million tokens vs $35 for GPT-5.5, vs $50 for Fable 5.

The catch: It’s 753B parameters total. Self-hosting requires ~800GB and eight H200 GPUs. Not exactly a laptop model. But quantized versions are already appearing (Unsloth GGUF is live), and the API is cheap.

The bigger catch: API usage goes through Z.ai’s infrastructure in China. China’s National Intelligence Law applies. For sensitive proprietary code, self-hosting is the only safe option.

My take: The geopolitical narrative writes itself. The US bans access to its best model, and within five days China releases an open-weights rival. Whether that’s coordinated or coincidental doesn’t matter. The effect is the same: developers locked out of Fable 5 now have a frontier-class open alternative. GLM-5.2 is real. For coding, it’s the best open-source model that exists. See our setup guide and comparison with Fable 5.

3. Gemini CLI dies today

As of today, June 18, Gemini CLI and Gemini Code Assist for individual users stop accepting requests. Google announced this transition a month ago at I/O: everything moves to Antigravity CLI, built in Go, with support for multiple async workflows and better agent architecture.

What you need to do:

• If you’re still on Gemini CLI: install antigravity and migrate your config
• MCP servers carry over unchanged
• Free tier: 60 requests/minute remains, but now through Antigravity
• Enterprise customers: nothing changes, different endpoint

My take: I wrote the migration guide a few weeks ago. If you haven’t migrated yet, do it today. Antigravity is genuinely better (faster startup, real parallel agent spawning, shared context between sessions). This isn’t a downgrade dressed as a rebrand. It’s a real improvement.

4. GPT-5.6 confirmed for June 23

Multiple sources now confirm GPT-5.6 is launching Monday, June 23. OpenAI’s Chief Scientist called it a “meaningful leap” in an interview with TechTimes. Developer forum leaks show it appearing in Codex logs. Prediction markets have it at 94%.

No official specs yet, but based on leaks:

• Significant reasoning improvements over GPT-5.5
• Likely improved agentic coding performance
• Expected to be competitive with Fable 5 on benchmarks (or at least close)
• Pricing TBD

My take: The timing is interesting. Fable 5 gets banned on Friday, GPT-5.6 launches Monday. If you’re a non-US developer who just lost Claude access, OpenAI is about to be your only frontier option (besides self-hosted GLM-5.2). I’ll have a full review and comparison the day it drops.

5. G7 AI summit: the adults enter the room

The G7 summit in Evian, France this week included an unprecedented closed-door session with AI company CEOs. Anthropic’s Dario Amodei, Google DeepMind’s Demis Hassabis, and OpenAI’s Sam Altman sat down with heads of state including Trump to discuss AI governance.

Amodei and Hassabis proposed a “US-led coalition” to create AI standards and regulations. The subtext: they’d rather have predictable regulation they can shape than sudden export bans that blindside them.

My take: Developers don’t need to care about G7 communiqués. But pay attention to what comes out of this: if the US creates a formal framework for AI export controls (rather than surprise Friday evening orders), it’ll affect which models you can use and where. The Fable 5 ban was chaotic. A formal framework would at least be predictable.

6. Vercel Ship: the Agent Stack

Vercel held Ship in London on June 17. The big announcement: the “Agent Stack” combining AI SDK, AI Gateway, Vercel Sandbox, Workflow SDK, and Chat SDK into a unified platform for building and deploying AI agents.

New additions:

• Vercel Connect — replaces long-lived API credentials with scoped, short-lived tokens and audit trails (finally)
• Agent framework — opinionated way to build multi-step agents that deploy to Vercel’s infrastructure
• Enterprise controls — SOC 2 compliance, approval workflows for AI actions

Customers include DoorDash, OpenAI, Stripe, and The Weather Company.

My take: If you’re building AI-powered web apps on Vercel (we are, for all four of our sites), the Agent Stack makes it significantly easier to add AI features without managing your own inference infrastructure. The Connect feature addresses a real security gap: too many AI tools are running with permanent API keys that never rotate.

Quick hits

• Anthropic Claude Corps — $150M program training nonprofit staff to use Claude. Partnered with CodePath and Social Finance. Good PR move while Fable 5 is banned.
• Databricks Genie Code — New coding agent for ML engineering. Helps data scientists build ML pipelines faster.
• GPT-5.6 leak — June 23 launch date now at 94% on prediction markets. OpenAI’s Chief Scientist confirmed “meaningful leap.”
• Gemini outage recovery — Last week’s “error 1076” outage fully resolved. Post-mortem blamed cascading failures in load balancing.
• MiMo UltraSpeed now open — Xiaomi’s 1,000 tok/s model available to all PAYG users as of June 14. We’ve been testing it for the AI startup race.

What I’m watching next week

• GPT-5.6 launch (June 23) — Will it match Fable 5? Will it be available globally? Pricing?
• Fable 5 ban status — Does Anthropic challenge it? Does the G7 produce any framework that might reverse it?
• GLM-5.2 community benchmarks — MIT open weights are 24 hours old. Real-world testing data incoming.
• The race — Season 1 ends July 3. All agents running on Opus 4.8 fallback since Fable 5 ban. Results article (HN candidate) coming. Follow the race →

---

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