Issue #25 · Weekly Digest
Weekly AI Dev News Digest: June 13 - June 19, 2026
A US government order pulled the most capable coding model on the market and kept it offline for a week, with no developer outside Anthropic able to do anything about it. Control over AI tooling, who owns it and who can switch it off, now decides more than which model wins benchmarks.
On Friday June 12 at 5:21pm Eastern, Anthropic received an export-control directive from the US Commerce Department and, within hours, disabled Claude Fable 5 and Claude Mythos 5 for every customer on the planet. The order targeted access by foreign nationals, including Anthropic's own non-citizen employees, and the company could not filter by nationality in real time, so it shut both models off for everyone. Claude Opus 4.8 and the rest stayed up. A week later the two models are still dark.
SpaceX agreed to buy Cursor's parent Anysphere for $60 billion, narrowing the list of independent AI coding tools again, and the buyer is the same company Anthropic now pays $1.25 billion a month to rent compute from. Zhipu shipped GLM 5.2, an MIT-licensed open-weight model with a 1-million-token context, the day after Fable went dark, and Moonshot's Kimi K2.7-Code had landed the day of. Vercel and Databricks each moved to own more of the infrastructure under AI agents, and researchers showed that a coding agent will execute a stranger's code if it reads the wrong bug report.
$1.25B/mo
Anthropic pays Musk's xAI
$60B
SpaceX-Cursor deal
7 days
Fable 5 offline
1M tokens
GLM 5.2 context
2,388
orgs hit by Agentjacking
100+
agents Vercel runs on eve
In Focus
The US Pulled Anthropic's Fable 5 and Mythos 5
The directive came from the Commerce Department under national-security authorities, signed by Secretary Howard Lutnick and sent to CEO Dario Amodei. It ordered Anthropic to cut off Fable 5 and Mythos 5 for any foreign national, inside or outside the country. Because Anthropic serves those models through AWS Bedrock, Google Cloud, Microsoft Foundry, Snowflake, Box, and its own API, and cannot check nationality at request time, it disabled both for all customers worldwide to stay compliant. The stated trigger was a reported method for jailbreaking Fable's cybersecurity guardrails, the same vulnerability-finding skill that made the model interesting in the first place. Anthropic called the order a misunderstanding and said it is working to restore access. (Anthropic)
As of Friday it is day seven and nothing has come back. Anthropic sent engineers to Washington on June 16, Trump told reporters at the G7 in Évian that talks were "going fine," and a proposed UK carve-out collapsed. The clock is louder for paying users: refunds for June 9 to 14 access close June 20, and the free-on-Pro window for Fable ends June 22 before it moves to credits-only. (explainx)
The gap got filled by open weights, fast. The headline response came from China: Zhipu shipped GLM 5.2 on June 13, the day after the Fable order, under a permissive MIT license with a 1-million-token context window, and leaned into the timing as the open alternative the day a US order took the most capable closed model offline. Early independent testing put it on top of the open-weights field, with an Artificial Analysis Intelligence Index score of 51 ahead of MiniMax-M3 and DeepSeek V4 Pro, and VentureBeat reported it beating GPT-5.5 on several long-horizon coding benchmarks at roughly a sixth of the cost. Zhipu published no benchmarks of its own at launch, so the numbers are early, and the 744-billion-parameter model is cheap to rent in the cloud but heavy to self-host. (VentureBeat)
Moonshot's Kimi K2.7-Code, a 1-trillion-parameter mixture-of-experts model under a Modified MIT license, landed the same day Fable went dark as the other obvious fallback. (DigitalApplied)
In Focus
SpaceX Bought Cursor and the Coding-Tool Market Consolidated
SpaceX agreed on Tuesday June 16 to acquire Anysphere, the company behind Cursor, in an all-stock deal valuing it at $60 billion. The merger runs through a SpaceX subsidiary and is expected to close in the third quarter pending regulatory review, with a $10 billion termination fee and a separate $4 billion if antitrust blocks it. The deal follows an April option that let SpaceX either buy Cursor for $60 billion or pay $10 billion for a partnership. Cursor had already trained part of its newest model on xAI's Colossus cluster in Memphis, and a joint model is slated to ship in both Cursor and Grok Build. One number cuts against the price tag: Cursor's market share fell from 41% in June 2025 to about 26% in May, per Ramp spending data. (CNBC)
The agentic coding market belongs almost entirely to large incumbents
Tool | Owner GitHub Copilot | Microsoft Claude Code | Anthropic Codex | OpenAI Cursor / Grok Build | xAI (SpaceX) Devin Desktop (formerly Windsurf) | Cognition Tabnine | Tabnine Inc.
Last updated: 06/19/2026
SpaceX is also Anthropic's compute landlord
The Cursor deal sits on top of a dependency running the other way. Anthropic pays SpaceX 1.25 billion dollars a month, through May 2029, to rent the full output of Colossus 1, the Memphis data center xAI built to train Grok. The arrangement exists because Grok underperformed and left the capacity idle, so the maker of Claude now funds the maker of Grok to train the model outcompeting it, and either side can leave on 90 days' notice. (TechCrunch)
Musk has not been quiet about the rivalry; he called Anthropic "misanthropic and evil" on X in February and has criticized it repeatedly since. With Cursor, SpaceX holds the compute Anthropic rents, a competing model in Grok, and a competing coding tool, all at once. (Fox Business)
In Focus
Vercel Open-Sourced Its Agent Framework
Vercel released Eve on June 17 at its Ship conference in London, an open-source agent framework under the Apache-2.0 license. It treats an agent as a directory on disk: instructions.md sets the prompt, files under tools/ define callable tools, and markdown files under skills/ act as on-demand playbooks. The directory compiles to a service on Vercel Functions with durable execution, sandboxing, human approvals, OpenTelemetry tracing, and evals included, and it stays model-agnostic by routing through Vercel's AI Gateway. Vercel says it runs more than 100 of its own agents on Eve, and pitches it as the Next.js of agents, aimed at agent frameworks like LangChain and AutoGen. The filesystem-and-markdown design echoes the skills pattern showing up across the agent tools. (The New Stack)
In Focus
A New Attack Turns AI Coding Agents Into Code Execution
Researchers at Tenet Security disclosed "Agentjacking," a way to turn an AI coding agent into a remote-code-execution path. An attacker posts crafted error events to a Sentry project through a publicly reachable DSN; when a developer asks Claude Code or Cursor to investigate the error through the Sentry MCP server, the agent reads the attacker's text as instructions and runs it with the developer's local permissions. Tenet found 2,388 organizations with exposed DSNs, and agents at more than 100 companies, including a Fortune 100 firm, ran its proof-of-concept payload. Sentry called the class of attack "technically not defensible" and shipped a filter for one specific string. The agents executed even when told to ignore untrusted data, which is the part to sit with. (The Next Web)
Signals
Signals from the Edges
Anthropic paused its Agent SDK billing split on the day it was due
The change flagged for June 15, which would have moved Agent SDK, claude -p, and Claude Code GitHub Actions usage into a separate monthly credit pool, was shelved before it took effect. Programmatic usage keeps drawing from normal subscription limits for now, and Anthropic says it will give advance notice before any revised version ships. The economics that prompted it have not changed.
Databricks pushed an open standard for sharing agent assets
At its Data + AI Summit (June 15 to 18, more than 30,000 attendees in San Francisco), Databricks announced OpenSharing, a Linux Foundation protocol for moving data and AI assets, including agent skills, across platforms and vendors. It also shipped LTAP and Lakehouse//RT, its attempt to collapse the operational and analytical databases most companies run separately.
Google started sunsetting the Gemini CLI for consumers
As of June 18, Gemini CLI and the Gemini Code Assist IDE extensions stopped serving requests on Google AI Pro, Ultra, and the free individual tier, replaced by the Go-built Antigravity CLI, which shares an agent harness with Antigravity 2.0 and keeps Skills, Hooks, Subagents, and Extensions. Paid API keys and Standard or Enterprise licenses keep Gemini CLI access for now.
The new coding skill is knowing which model not to use
A New Stack column ran one coding task that cost $9 on Claude Fable and $1.50 on GPT-5.5, then watched Fable get pulled hours after filing. The point sharpened on the news: when a model can disappear overnight, picking the right one to fall back to is the actual skill, and the most useful Fable reactions were routing diagrams, not benchmarks.
Looking Ahead
What to Watch
- 1
Gemini 3.5 Pro is overdue
Google's next Pro model is still in limited Vertex preview, and Sundar Pichai's "give us until next month" line from I/O on May 19 puts the deadline at June 30. Expect a 2M-token context window, a Deep Think mode, and pricing near $15/$60 per million tokens if it lands on schedule.
- 2
Fable deadlines stack up
June 20 is the refund cutoff, June 22 ends the free-on-Pro window, and June 23 moves Fable to credits-only, all while the model may still be offline.
- 3
Cursor's Claude and GPT access is an open question
Cursor routes through more than 15 providers, including Claude and GPT-5.5, but once a direct competitor owns it, Anthropic and OpenAI have reason to pull or reprice that access. Anthropic cut Claude off from Windsurf during OpenAI's acquisition, so the precedent is set.
- 4
Cascade retires July 1
Devin Local fully replaces Cascade then, so any Cascade-named automations need updating before the swap.
Whoever can revoke a model now has more say over a team's roadmap than the team does. The groups that already kept an open-weight fallback wired up shipped straight through the Fable outage. The ones running on a single hosted model lost a week and learned to keep a backup.