AI Bulletin · week of June 15

This week’s bulletin gathers six news blocks on AI applied to work, from 15 to 19 June 2026, plus a closing framework on model sovereignty. It covers the withdrawal of two Anthropic models by order of the US government, a batch of model launches and announcements, a large funding round in China and several moves in agents, infrastructure and chips.

1. The US government forces the withdrawal of Claude Fable 5 and Mythos 5

Anthropic suspended access to Claude Fable 5 and Claude Mythos 5 for all users after receiving a directive from the US government. The company linked the measure to export controls, national-security concerns and reported jailbreak risks.

According to this week’s reports, conversations between Amazon’s CEO, Andy Jassy, and US officials prompted the Administration to halt all foreign use of Anthropic’s most capable models. Amazon researchers had used a series of prompts to get Fable 5 to provide information that could be used to aid cyberattacks, and White House officials asked Anthropic to fix the vulnerabilities or take down the model.

Anthropic shut down access to Mythos and Fable to comply, although it maintains that the vulnerabilities flagged by Amazon are relatively basic and that other publicly available models are also capable of discovering them. A subsequent experiment with 26 combinations of Anthropic and OpenAI models found that more reasoning effort, and even more recent models, do not always improve the detection of that kind of security flaw.

In parallel, Anthropic paused the token-based billing change it had announced the previous month for its Claude Agent SDK, just before it was due to take effect. Usage outside the SDK will be billed at the company’s prevailing API rates.

Sources: Ars Technica — “Anthropic pauses token-based billing for its Claude Agent SDK”; reports via TLDR AI, 15 and 18 June — “Amazon CEO’s talks with US officials triggered crackdown on Anthropic models” and “Brain the Size of a Planet: Are LLMs Thonking Too Hard?”.

The bulletin mascot as a reporter holding up a blank document sealed with a large red wax stamp, with the empty facade of a columned government building visible through the window — The US government forced Anthropic to suspend access to Fable 5 and Mythos 5 for all users.

2. Models: GLM-5.2, Kimi K2.7 Code and the GPT-5.6, Cursor and Mistral announcements

Z.ai launched GLM-5.2, its new flagship model, with a one-million-token context window, new reasoning controls and support for long-horizon coding tasks across entire codebases. The model is available immediately to GLM Coding Plan users, with API access, chatbot support and MIT-licensed open weights planned for the following week.

Moonshot released Kimi K2.7 Code, an agentic coding model with a mixture-of-experts architecture and one trillion total parameters, accessible through its OpenAI- and Anthropic-compatible API. An experiment that generated twelve landing pages with each tool put Kimi K2.7 Code 94% lower in cost, sixteen times cheaper than Claude Fable 5.

OpenAI is preparing to launch GPT-5.6, possibly with Mini and Pro variants, the following week, with a 1.5-million-token context window, improvements in long-horizon coding and faster Codex responses, according to TestingCatalog. Added to that are two announcements with no fixed launch date: Cursor previewed a model of its own trained from scratch on more than 100,000 GPUs and with more than 1.5 trillion parameters, expected in the coming weeks, and Mistral confirmed a new model for this summer, with an early-access programme in July for research, government and industry partners.

Sources: Z.ai — “GLM-5.2”; Hugging Face — “Kimi K2.7 Code”; TestingCatalog — “OpenAI prepares GPT-5.6 models for the upcoming release”; reports via TLDR AI, 18 June — “Cursor’s New Model”, “Mistral Has a New Model Coming This Summer” and “Kimi K2.7 Code vs Claude Fable 5”.

3. DeepSeek raises $7.4 billion and becomes China’s most valuable AI startup

DeepSeek raised more than $7.4 billion in a round that valued the company at more than $50 billion, making it China’s most valuable AI startup. Its founder, Liang Wenfeng, invested around $3 billion in the deal; before the round he controlled close to 90% of the company. A government-backed fund contributed around $150 million.

The company will use the new capital for R&D and to expand its computing infrastructure. In parallel, according to CNBC, DeepSeek asked potential investors not to poach its staff or encourage them to start their own companies.

Sources: CNBC — “‘No poaching’ our people, China’s AI behemoth DeepSeek reportedly tells investors”; report via TLDR AI, 17 June — “DeepSeek becomes China’s most valuable AI startup after $7.4 billion fundraise”.

4. Agents and coding: Cursor Origin, Vercel, Claude Code, Codex and Replit

Cursor unveiled Cursor Origin, a GitHub competitor built for agents: a Git-compatible forge built on the assumption that numerous agents will clone, branch, commit, rebase and review in parallel. The company frames it as a piece of an agent-operated software development chain, as opposed to a GitHub designed for human-scale development.

Vercel launched two products. Connect, in public beta, replaces long-lived provider tokens with a runtime credential exchange that issues short-lived, task-scoped credentials for agents. eve is an open-source agent framework with durable execution, sandboxed compute, approvals, subagents and evaluations built in.

Anthropic added support for “artifacts” to Claude Code, turning work sessions into shareable visual pages, with version history and privacy controls, in beta for Claude Team and Enterprise. OpenAI added support for the Chrome DevTools Protocol in Codex, available through opt-in settings and excluded in the EEA, the UK and Switzerland, and Replit became integrated within Claude.

Sources: Vercel — “Introducing Vercel Connect” and “Production infrastructure for AI agents”; Claude — “Artifacts in Claude Code”; TestingCatalog — “OpenAI released CDP support for browser use on Codex”; Replit — “Replit is now available in Claude”; report via TLDR AI, 17 June — “Cursor Origin”.

The bulletin mascot at an oak newsroom desk directing several tiny faceless mechanical arms that assemble and stamp blank papers at miniature workbenches, under a studio lamp — The week brought a batch of agent and coding tools: Cursor Origin, Vercel's Connect and eve, Claude Code artifacts and browser support in Codex.

ChatGPT’s market share fell below 50% for the first time, according to TechCrunch, with users migrating to other assistants such as Google’s Gemini, Anthropic’s Claude and xAI’s Grok. ChatGPT was the fastest app ever to reach a billion monthly users and remains the most popular AI assistant in the world; the figure points to users’ greater willingness to switch between assistants.

OpenAI also improved ChatGPT’s scheduled-tasks feature, now accessible from a new “Scheduled” page for the Go, Plus, Pro, Business and Enterprise plans, and retired Pulse.

Meta launched an “AI Mode” on Facebook that turns the search bar into a conversational tool answering queries from public group posts, Reels and Marketplace, rolling out in the US. Google introduced new agent capabilities in Android 17 with AppFunctions and Android MCP, which let apps expose tools that on-device agents can discover and execute. And Perplexity unveiled Brain, a memory system that builds a persistent context graph across tasks, projects, decisions, files and sources.

Sources: TechCrunch — “ChatGPT’s market share slips below 50% for first time”; Android Headlines — “Facebook gets its own AI Mode”; Android Developers Blog — “Android 17”; Perplexity — “Self-improving memory for agents”; report via TLDR AI, 18 June — “ChatGPT improves scheduled tasks and retires Pulse”.

6. Infrastructure, chips and hires: NVIDIA, Google, Qualcomm and Noam Shazeer

NVIDIA’s Blackwell platform dominated the MLPerf Training 6.0 benchmarks, with the fastest training times and the largest-scale training, on 8,192 GPUs. In a separate measurement this week, the Blackwell Ultra NVL72 platform topped the AgentPerf benchmark with twenty times more agent throughput per megawatt than the Hopper generation.

Google is renting compute capacity from its processors to Anthropic at an AI data centre in western New York State, a move that, according to reports, helps those data centres raise cheaper debt and that the company is using to build a chip business to rival NVIDIA’s. Qualcomm announced the Snapdragon Reality Elite platform for mixed-reality glasses and a “Scalable Turnkey AI-Ready” kit, as part of work on more than 40 AI wearable devices.

In personnel moves, Noam Shazeer, co-author of the 2017 paper “Attention Is All You Need”, is leaving Google to join OpenAI.

Sources: NVIDIA — “Fastest, Largest, Strongest: NVIDIA Blackwell sweeps MLPerf Training 6.0” and “NVIDIA Blackwell leads on first agentic AI infrastructure benchmark”; TechCrunch — “Qualcomm wants to be the chip inside whatever replaces your smartphone”; reports via TLDR AI, 18 and 19 June — “Noam Shazeer is joining OpenAI” and “Google is using NVIDIA’s playbook to build a rival AI chip business”.

The bulletin mascot inspecting a glowing silicon chip wafer next to a server rack and a stack of wafers on an oak table, with a coffee cup and a monstera plant — A week loaded with infrastructure: NVIDIA swept MLPerf Training 6.0, Google rented chip capacity to Anthropic and Qualcomm expanded its work on AI devices.

7. Framework of the week: model sovereignty and “owning vs renting” intelligence

The withdrawal of Fable 5 and Mythos 5 by government order gave rise to several pieces that circulated this week on who controls the models companies use.

“Owning vs. Renting Intelligence”, by Fireworks’ CEO, Lin Qiao, argues that a company built on intelligence it does not own found itself exposed to decisions it could not influence, and cites the case of tuned open models that match frontier quality at a fraction of the cost.
“Should You Post-Train Your Own Model?” holds that, for the handful of use cases critical to a company’s product and margin, the answer is increasingly to post-train a model of one’s own rather than depend on a general model.
“Sovereign AI Is Not a Model, But a Supply Chain Problem” frames sovereign AI as the question of how much of the supply chain needed to train, operate, validate and protect the models can be secured within one’s own country or allied nations.
Two analyses of the government’s decision, “Leviathan Waking” and “That untravell’d world”, describe a new and more political phase of AI governance, with disputes still unresolved between the Administration and Anthropic.
Yann LeCun, carried by CNBC, called Elon Musk’s xAI a failure and warned that the labs must raise prices or cut costs to avoid what he called a “big bubble explosion” due to their high operating expenses.

Sources: CNBC — “Godfather of AI blasts Musk’s xAI as ‘failure’, says labs are risking a ‘big bubble explosion’”; Hyperdimensional — “Leviathan Waking” and “That untravell’d world”; reports via TLDR AI, 16 June — “Owning vs. Renting Intelligence”, “Should You Post-Train Your Own Model?” and “Sovereign AI Is Not a Model, But a Supply Chain Problem”.

Closing

The six blocks above cover the withdrawal of Claude Fable 5 and Mythos 5 by order of the US government and Anthropic’s pause on its Agent SDK billing, a batch of models with GLM-5.2, Kimi K2.7 Code and the GPT-5.6, Cursor and Mistral announcements, DeepSeek’s $7.4 billion round, the agent and coding launches from Cursor, Vercel, Anthropic, OpenAI and Replit, the consumer-assistant moves from OpenAI, Meta, Google and Perplexity, and a round of infrastructure, chips and hires involving NVIDIA, Google, Qualcomm and Noam Shazeer’s move to OpenAI. The framework of the week gathers five readings on model sovereignty that emerged in the wake of the government decision. The items reflect what the companies and the cited reports have communicated; they do not include productivity data verified in specific organisations.