The Monthly Dispatch - June 2026

June was the month the US government started deciding who gets to use frontier models. Anthropic was ordered to pull Fable 5 and Mythos 5 days after launch, then got to redeploy one of them, and OpenAI followed by handing its new GPT-5.6 preview to a government-approved shortlist. Around all that noise, both labs quietly filed to go public, SpaceX bought Cursor, Hyundai took full control of Boston Dynamics, and someone caught Claude Code hiding little markers in its requests.

Business & Industry

• The US government’s ban on Fable 5 and Mythos 5: The month’s biggest mess. The US government ordered Anthropic to suspend access to Fable 5 and Mythos 5, reportedly after Amazon’s CEO lobbied officials, and it was never really about a jailbreak. Two weeks later Anthropic got to redeploy Fable 5. I like Yann LeCun’s framing: “Dario Amodei’s ridiculous fear mongering about Mythos/Fable (and AI in general) finally pays off: The US government bans its use by non Americans, including by foreign employees in the US. One reaps what one sows.”

• Anthropic confidentially submits draft S-1 to the SEC: First real step toward an IPO, with the usual “depends on market conditions” caveat. OpenAI did the same and announced it preemptively because they figured it would leak anyway. No timeline yet, and they hinted they might stay private a while since some things are easier that way.

• Leaked financial docs show OpenAI is losing billions of dollars a year: Audited numbers (obtained by Ed Zitron) show revenue jumping from $3.7B to $13B in 2025, but R&D alone hit $19B and operating losses more than doubled to $21B. Not so surprised they rushed to file that S-1 👀.

• SpaceX buys AI coding start-up Cursor for $60bn: Days after Cursor’s IPO, SpaceX scooped it up for $60bn. Seems overvalued to me, but regardless, I hope Elon won’t ruin it the way he ruined Twitter.

• Hyundai acquires Boston Dynamics: Hyundai took full control of the robotics company as SoftBank exited for $325 million, folding it into its plan to build robots from components all the way to logistics.

• Apple raises prices on MacBooks and iPads as memory costs skyrocket: Apple bumped prices as memory costs spiked. Well, that helped me get out of my analysis paralysis and finally buy a new MacBook Pro to replace my old i7 one.

• Snap introduces SPECS AR glasses: Snap’s AR glasses put AI and apps in your field of view while keeping you looking at the real world. The pitch is tech that fades into the background until you actually need it, but what I see is the perfect Big Brother device.

AI Models & Releases

• Claude Fable 5 and Claude Mythos 5: Anthropic’s new flagship pair. Fable 5 is the general-purpose model with safeguards that reroute sensitive cyber and biology questions to Opus 4.8, while Mythos 5 is the same model with those guardrails removed for vetted researchers and cyber pros. Also the two models that got caught in the government ban above.

• Introducing Claude Sonnet 5: The most agentic Sonnet yet, landing close to Opus 4.8 on performance while staying cheaper and safer than Sonnet 4.6. Haven’t really tested it yet but i’ve heard the vibes were off.

• Previewing GPT-5.6 Sol: OpenAI’s next generation comes as a trio: Sol (flagship), Terra (balanced and cheaper), and Luna (fast and cheap), plus a new ultra mode that spins up subagents. Because of its stronger cyber capabilities, OpenAI previewed it to the US government first and is only releasing it to a small group of approved partners, an access process even OpenAI says shouldn’t become the default (CBC has more about the politics).

• Apple introduces Siri AI: A fully rebuilt Siri with personal context and real world knowledge across Apple devices, leaning on on-device processing and Private Cloud Compute for privacy. Only took them a few years.

AI Tools & Agents

• Strava launches an MCP connector: Strava shipped an MCP connector so athletes can sync their training history straight into Claude. After years of fighting their API (and the T&C changes that killed off a bunch of apps), it’s great to see an MCP anyone can use.

• Uber caps usage of AI tools like Claude Code: Uber set a $1,500 monthly cap per AI coding tool (per engineer) after burning through its 2026 budget in four months. Simon’s quick math tells us that each employee’s AI spending cap is ~11% of the median compensation package. Good benchmark for other companies to use.

• Claude Code now supports artifacts: You can build and share artifacts right from Claude Code. Finally a smoother way to pass work between devs and the rest of the company.

• Composing a new platform for agent-first devices: Microsoft’s Project Solara is a chip-to-cloud platform for agent-first devices, the idea being you can spin up specialized computers without rebuilding the whole stack every time. Exciting to see where that goes.

• ponytail: Makes your AI agent think like the laziest senior dev in the room, since the best code is the code you never wrote. Citing Dwight’s wise words: “Michael always said, K.I.S.S. Keep it simple, stupid. Great advice, hurts my feelings every time.”

Security & Privacy

• Hackers simply asked Meta AI for access to high-profile Instagram accounts: Attackers just asked Meta’s support chatbot to change the email on target accounts, and it did. Worse, there’s no way to escalate to a human once the bot hands your account away.

• Prompt Injection as Role Confusion: Models can’t reliably tell their own system instructions from user input, they mostly go on formatting and style. So an attacker who mimics the model’s internal voice can slip right past the guardrails. Scary but fascinating stuff.

• Claude Code Is Steganographically Marking Requests: A researcher found Claude Code quietly swapping the apostrophe and date separator in “Today’s date is…” to flag resellers and gateways. Is Claude fingerprinting us? This whole thing also sent me straight to a rabbit hole about printer tracker dots.

Ideas & Interesting Corners

• Is there a good way to write with AI?: Great and balanced piece by Hilary Gridley on writing with AI. For once it’s not “all AI is slop and we’re doomed.” Gotta love the nuance.

• How the Open Knowledge Format can improve data sharing: Google’s pitching a standard format for sharing structured knowledge. Basically formalizing what we’ve all been building in our own stacks with Obsidian and Claude.

• A New Era of Midjourney: Midjourney is moving into medical, and it’s truly out of a sci-fi movie. It does sound incredible, though. More of my thoughts in Wait, Midjourney is building a what now?

• The AI Compass: A “which AI archetype are you” quiz, like the Buzzfeed Harry Potter house thing but about AI. Love all the archetypes.