Anthropic disables Fable 5 and Mythos 5 as SpaceX moves to buy Cursor

Anthropic disabled Fable 5 and Mythos 5 for all customers worldwide after receiving a US Commerce Department export-control directive on June 12, 2026.¹² According to the company, the directive required it to suspend access for foreign nationals, including those inside the United States.³² It was not an outright prohibition on the models for every customer; the worldwide shutdown was Anthropic’s response to the access restriction.³¹

US officials cited national security and cybersecurity concerns.⁴³ Anthropic said officials had not publicly provided specifics and that, to its knowledge, the concern included an alleged technique for bypassing Fable 5’s safeguards.⁴³ After reviewing a demonstration, Anthropic said it found only a small number of previously known, minor vulnerabilities. It said similar issues could be found using other publicly available models.⁴

On June 16, 2026, SpaceX agreed to acquire Anysphere, the company behind the Cursor coding tool.¹² The all-stock deal values Anysphere at $60 billion, but the acquisition has not closed.¹² Cursor is expected to become a SpaceX subsidiary when the transaction is completed.³²

The agreement follows a previously negotiated purchase option. It would give SpaceX a larger role in AI-assisted software development, a market where Anthropic and OpenAI are also active.¹³ The reports do not confirm a jointly developed coding model or a specific plan to use SpaceX computing infrastructure for one.¹³ For companies, the potential change in ownership is more concrete than any new model platform.

Z.ai introduced GLM-5.2 for long-running coding tasks and released the model weights on June 17. The company states that the model supports a context window of up to one million tokens and positions it for work on large software projects.¹² GLM-5.2 is also available through an API.³

The large context window could let teams work with more code and documentation in a single task. Its availability for every use case needs checking, however: a separately named variant is associated with the one-million-token window.⁴ Z.ai reports improvements on coding tasks, but results depend on the evaluation used.¹ The released weights also give companies an option to assess self-hosting, subject to compute requirements and license terms.¹⁵

Google introduced AppFunctions on June 16, 2026, as a way for AI agents to discover and execute Android app capabilities on a user's behalf.¹² Agents can access an app's local state directly to complete workflows within it.¹ The accompanying Jetpack library launched in alpha, while Gemini integration is in private preview with trusted testers.¹²

For now, AppFunctions is a developer platform capability, not a broadly available way to control any app through Gemini.¹² Companies considering support will need to decide which actions agents may perform and which local data those actions expose. Google offers an early access program to help developers prepare their apps, but has not said when Gemini integration will become more widely available.²

On June 18, 2026, Adobe expanded its Creative Agent across Firefly and Creative Cloud.¹ Its AI Assistant is available in public beta in Premiere, Photoshop, Illustrator, Frame.io, and InDesign.²¹ Adobe says it can coordinate multistep tasks through natural-language prompts, including organizing footage, renaming assets, and adjusting layouts.²

The expansion puts more emphasis on managing production workflows rather than generating individual assets alone.²¹ Adobe also plans to make its creative tools accessible through ChatGPT, Claude, Copilot, Gemini, and Slack, but the integrations are not all available at once.¹³ New assistant capabilities in the Firefly web app are available in beta.¹ Adobe is also testing an updated studio workspace with limited initial access.¹

Apple announced Siri AI and a dedicated Siri app at WWDC on June 8.¹² Developers can test the new features on iOS 27, iPadOS 27, macOS 27, and visionOS 27; a user beta is planned for later in 2026.¹³ Siri is designed to resume conversations and answer questions using personal context from messages, email, photos, and on-screen content.⁴¹

Siri AI is also intended to draw on web information and take actions in apps.⁵³ Depending on the task, Apple will process requests on-device or through its private cloud systems.⁴⁵ The overhaul is part of Apple's effort to catch up with competing AI assistants.⁵ Its reliability in everyday use remains to be tested in the broader beta. The announced app is not yet a generally released consumer product.¹²

Prometheus, led by Jeff Bezos and Vik Bajaj, announced a $12 billion funding round on June 11, 2026. Reports put its valuation at about $41 billion. The company aims to develop AI systems that help design and manufacture complex physical products, including jet engines and computers. Some of the funding is intended to expand computing capacity.¹²³

That ambition puts Prometheus in engineering workflows where design and manufacturing are closely linked. Bezos describes the goal as an “artificial general engineer” for the physical world. It is not yet possible to assess publicly how well its systems perform these tasks: The coverage provides no verified benchmarks or date when customers can use a product. For now, the round establishes the scale of the investment rather than the practical value of the planned systems.²⁴

Moonshot AI released Kimi K2.7 Code on June 12 as an open-source model for agentic coding. It is available through Kimi Code and the Kimi API. Cloudflare also added the model to Workers AI.¹²

Moonshot reports 30% lower reasoning-token use than K2.6 and a 21.8% higher score on Kimi Code Bench v2. Those figures come from the company; the available sources do not establish independent verification of the benchmark gains. Fewer reasoning tokens also do not necessarily mean an entire task runs 30% faster or costs 30% less. For companies, the release adds another option for coding agents, but its value depends on performance against their own tasks and repositories.³⁴

MiniMax released the weights for MiniMax M3 on Hugging Face on June 12, 2026.¹² The company describes M3 as accepting text, images, and video, and reports a one-million-token context window.¹ According to NVIDIA, the model uses MiniMax Sparse Attention, with a prefiltering stage intended to reduce the cost associated with conventional quadratic attention.³

The combination of published weights, multimodal input, and a large context window could be useful for applications handling extensive documents or video.¹³ The reported context length has not been independently verified. Public materials also do not establish which serving conditions allow use of the full window.¹³ MiniMax’s benchmark results offer a starting point for evaluating performance, but teams should test the model on their own workloads.¹

In mid-June, LMCache published technical details about its already available open-source KV-cache layer. A June 15 post explains how cache data moves in multiprocess deployments.¹ On June 17, the project also presented an access path using GPUDirect Storage.² LMCache works with inference engines such as vLLM to reuse stored intermediate results across requests.³

That reuse can shorten time to first token, especially when requests share substantial context. LMCache says the improvement can reach an order of magnitude in high-reuse scenarios. This is a conditional project claim, not a benchmark result that applies to every deployment.³ For operators, the practical benefit depends on how often content is requested again and which storage and transfer paths their infrastructure provides.³¹