The Brief.

Five days, four moves, one story: the lock-in stopped being the model. Apple paid Google for the frontier and kept the device. OpenAI bought the runtime, the EU finished the labelling code, and Washington started arguing about who owns the upside. The model is the easy decision now — everything around it is where the next two years live.

Anthropic shipped Claude Fable 5 today — the first publicly available Mythos-class model, the same architecture they spent the last two months calling too dangerous to release broadly. The bridge between "too dangerous" and "public" is a runtime classifier that routes prompts in cybersecurity, bio/chem, and model-distillation categories to Claude Opus 4.8 instead of Fable. Anthropic reports the classifier fires on under 5% of sessions. The user never sees a refusal; they get a less capable answer from a real model. This is the deployment pattern I have been recommending to portfolio teams in regulated verticals for about eighteen months. It is good to see a frontier lab ship the cleanest public version of it. Three things to copy if you build a version yourself: invest in the classifier as a product in its own right, pick a fallback that can stand on its own, and route per prompt, not per session.

Moonshot dropped Kimi K2.7-Code on Friday. The headline number: open weights, 32B active out of 1T, and about $1/$4 per million tokens versus roughly $15/$75 for Claude Opus 4.8. On Moonshot's own MCPMark Verified run, K2.7-Code beats Opus 4.8 (81.1 vs 76.4) but trails GPT-5.5 (92.9). It is a coding-and-agent model, not a general one. Two takeaways. First, the open-source floor for agentic coding just moved up by a meaningful step; if you build coding agents, re-run your top-50 traces against it this week. Second, the benchmark is the vendor's own, so treat the deltas as a hypothesis until a third party reproduces. The cost gap is real either way. Where it matters most is the long-tail traces where you were burning frontier prices on routine work.

Apple shipped a multi-model AI picker today at WWDC. iOS 27 will let users set Claude, Gemini, or ChatGPT as the default model behind Siri, Writing Tools, and Image Playground. Developers get an Extensions API to plug in their own. The architecture is what matters. Siri now routes in three tiers: on-device for simple tasks, Apple Private Cloud Compute in the middle, Google Cloud on Blackwell B200s for hard reasoning. Each tier has its own privacy boundary and cost profile. That is policy-gated routing in the wild, on the biggest consumer surface there is. Apple pays Google roughly $1B a year for the heavy tier. The next twelve months of pricing at the model layer will run through how Anthropic, OpenAI, and others negotiate into a slot where the user, not the integrator, is the customer.

Apple's Xcode 27 ships a LanguageModel protocol that lets developers swap on-device, Claude, Gemini, or GPT without changing application code. The interesting part is not the abstraction; it is the side effect. Provider lock-in for any Apple-platform app is now a configuration choice, not an architecture choice. The coding agents inside Xcode 27 also get the tools to validate their own work: write tests, run them, check previews, drive the simulator. That loop pattern is what production agent teams have been building toward for a year. Apple put it in the IDE. Anthropic published a matching Swift package the same day. The right metric for an agent-in-IDE is not lines of code generated; it is the rate at which you accept its diff without touching it. Track that one.

The European Commission shipped the final Code of Practice on labelling AI-generated content on June 10, ahead of Article 50 of the AI Act going live August 2. The code is voluntary; it gives signatories a presumption of conformity. The penalty floor for non-compliance with Article 50 is up to EUR 15 million or 3% of global annual turnover. What is actually in it: metadata, watermarking, and visible indicators for AI-generated images, audio, video including deepfakes, and text on matters of public interest, plus a duty to tell users when they are talking to a chatbot. Two things I would do this month. Audit every modality your product generates and confirm a watermarking spec covers it. Then decide whether you sign the code — supervisory authority conversations are where the presumption pays. The teams that treat August 2 as the deadline rather than the start of the conversation are the ones who will pay for it.

Anthropic ended zero-data-retention agreements for Mythos-class traffic today. Fable 5 and Mythos 5 traffic is retained for 30 days. The retention applies even to enterprise customers who signed ZDR contracts last quarter, and even to traffic routed through AWS Bedrock, Google Cloud, or Microsoft Foundry with ZDR turned on. Anthropic says the data is used only for safety work — defending against novel jailbreaks and reducing false positives — and never for training. I believe them. That is not the question. The question is what your compliance team does with it. If your customer contracts promise zero-retention upstream, you have a contractual delta between what you sold and what you ship. The remediation is meetings, not engineering. For most regulated workloads the right answer this quarter is probably to stay on Opus 4.8 until contracts catch up.

The Trump administration is floating direct federal equity in leading AI labs. OpenAI's preferred path, sketched in its April policy paper, is to donate shares into a 'Public Wealth Fund.' No cash from taxpayers, no licensing regime, no formal regulator. Just an ownership relationship between the labs and the state. Anthropic is not in talks, consistent with February's split after the lab refused to drop guardrails for Pentagon use. The fork is real: labs that take the equity deal and trade independence for procurement access, and labs that hold the line. Watch the enterprise RFP language over the next two quarters. 'Government does not have a board seat' is going to start showing up as a buying criterion in regulated domains.

OpenAI is buying the sandbox. On June 11 OpenAI agreed to acquire Ona, the Kiel-based cloud startup formerly known as Gitpod. Ona gives AI agents pre-configured cloud environments to actually finish work; the team folds into Codex, which OpenAI now puts at five million weekly users. The interesting datum is not the deal, it is that number. Five million weekly is not a beta. The pattern across the last month — agent commerce protocols, the cloud rebuilds at the labs — is that frontier labs are no longer competing on model quality alone. They are competing on the surface the model lands on. If your AI roadmap still ends at 'we picked a model', you are planning at the wrong altitude. Watch how many of OpenAI's enterprise pilots end up running on Ona environments inside the year. The model is the easy decision. The runtime is the lock-in.

OpenAI launched a Deployment Company today, in partnership with 19 investment firms, consultancies, and system integrators led by TPG, with Advent, Bain Capital, and Brookfield as co-lead founding partners. They are also acquiring Tomoro, an applied-AI engineering shop. Read this alongside Anthropic's Services Track and Partner Hub, formalized last week. The pattern is clear: the labs are no longer waiting for the systems integrators to build the implementation layer. They are buying or partnering into it. If you sell AI consulting, the channel just got more crowded. If you buy AI consulting, the question to ask now is which lab's deployment company your SI is in. If you are a startup planning to wedge into 'help enterprises ship Claude or GPT,' the wedge just got narrower. The most interesting move is the one neither company has announced yet: who owns the eval framework.

Sam Altman, Demis Hassabis, and Dario Amodei are all attending the G7 summit in Évian-les-Bains next week, June 15-17. It is the first G7 with all three frontier-lab CEOs in the room, and the first for Altman. None has signalled what they plan to discuss, which is a tell on its own. Two things to track. One, the timing: the EU AI Act's Article 50 enforcement begins August 2, and the GPAI Code of Practice on content labeling shipped this month. The labs have a narrow window to negotiate before enforcement lands. Two, who is not in the room: no Chinese lab, no Mistral, no open-weights camp. The geometry matters as much as the agenda. If you build under EU jurisdiction, the outcomes here will likely shape what your system cards have to look like by Q4.

If your AI roadmap still ends at 'which model', take the rest of June to rewrite it around the runtime, the eval, and the contract. Reply with what you are seeing in your own stack — I read everything.