A Frontier Model Launched, Then Got Shut Down by the US Government - PM Status Report, 15 June 2026
- Anthropic: Launched, Then Switched Off
- OpenAI: Agents That Can Pay and Keep Working Overnight
- Apple and Google: Siri Gets a Gemini Brain
- What This Means for Project Environments
On 9 June, Anthropic released Claude Fable 5 - the most capable publicly available model on several benchmarks. Three days later, the US Commerce Department ordered Anthropic to suspend it worldwide. Anthropic complied within hours, pulling both Fable 5 and its restricted sibling Mythos 5 for every customer, everywhere, with no notice period.
That’s the headline. But it wasn’t the only thing that happened. OpenAI partnered with Visa to let AI agents make payments under spending limits. Apple confirmed Siri will run on a custom Google Gemini model. OpenAI bought a cloud-execution company to keep coding agents running after you close your laptop. And Google shipped a text model that generates 256-token blocks in parallel instead of word by word.
Individually, each of these is a feature update. Together, they describe an industry moving fast enough that a model can go from launch to global shutdown in 72 hours - and that’s now a live consideration for anyone building workflows on top of these tools.
Anthropic: Launched, Then Switched Off
Fable 5 posted genuinely strong numbers, and Stripe reported it completed a 50-million-line codebase migration in a day that would otherwise have taken a team over two months. Mythos 5, the same model with safety classifiers lifted, went to vetted cybersecurity partners only.
On 12 June, Commerce issued an export-control directive citing a “narrow jailbreak” that let the model identify software vulnerabilities - a capability Anthropic says is no different from what GPT-5.5 can already do. Because the order applied to foreign nationals and Anthropic couldn’t segment its global user base on short notice, it disabled both models for everyone. Anthropic disagrees publicly with the decision; the White House has signalled it wants to lift the restriction once the underlying issue is patched, but as of 15 June, Fable 5 and Mythos 5 remain offline. All other Claude models, including Opus 4.8, are unaffected.
Separately, Anthropic landed enterprise alliances with TCS (50,000 employees across 56 countries) and DXC (banks, airlines, regulated sectors), plus an expansion of its Project Glasswing cybersecurity programme to roughly 150 organisations.
OpenAI: Agents That Can Pay and Keep Working Overnight
Two announcements here matter more than they might first look.
Visa integration (10 June). OpenAI’s agent platform can now execute payments through Visa’s network, with tokenised credentials, spend limits, approved merchant categories, and mandatory authorisation checks. This is the first mainstream “agent with a wallet” framework from a major lab - agents that can actually transact, not just recommend.
The Ona acquisition (11 June). OpenAI announced it’s acquiring Ona (formerly Gitpod), pending regulatory approval, bringing persistent, customer-controlled cloud execution environments into Codex. The practical effect: agents can keep running - accessing tools, databases, and credentials - after you’ve closed the laptop or ended the session. Codex passed 5 million weekly users this week, up from 3 million in April, and OpenAI says knowledge workers are now roughly 20% of that base and growing three times faster than developers.
Also this week: a simplified model picker (Instant/Medium/High/Extra High replacing the old Thinking variants), continued “Dreaming” memory rollout with better user controls (delete, edit, or turn off individual memories), and GPT-Rosalind - a life-sciences reasoning model now expanded with sponsored access for public health agencies doing biosecurity screening.
Apple and Google: Siri Gets a Gemini Brain
At WWDC on 8 June, Apple confirmed what had been rumoured for months: a rebuilt Siri running on a custom Google Gemini model, reportedly costing Apple around $1 billion a year. Routing is tiered - on-device first, then Apple’s Private Cloud Compute, then Google’s infrastructure for the heaviest reasoning. iOS 27 also lets users set Claude or ChatGPT as their default assistant for the first time. Apple’s software chief Craig Federighi was emphatic that this doesn’t change Apple’s privacy stance: requests aren’t stored or made accessible to Google.
Google also shipped Gemini 3.5 Live Translate - continuous, speech-to-speech audio translation (not turn-by-turn) with only a few seconds of latency, supporting 70+ languages and already live in Grab’s ride-hailing app and Google Meet. And on the research side, Google DeepMind released DiffusionGemma, an open 26B-parameter model that generates text in parallel blocks rather than one token at a time - 4-5x faster inference, useful anywhere low latency matters more than peak quality.
What This Means for Project Environments
Multi-model routing isn’t optional anymore - it’s risk management. Fable 5’s shutdown is the clearest demonstration yet that a frontier model can disappear overnight for reasons entirely outside a vendor’s control. If a workflow depends on one model from one provider, that’s now a documented single point of failure. Build fallback paths (Anthropic’s own safety routing falls back to Opus 4.8 by design) and avoid premium-model lock-in for anything mission-critical.
Agentic payments change what “approve and send” means. The Visa integration is early, but it points at a near-term world where an agent can book travel, pay a supplier invoice, or renew a subscription within limits you set - without a human clicking “confirm” each time. For project environments, this is worth getting ahead of: define spend thresholds and approval categories for agent-initiated transactions before someone else does it for you under pressure.
Persistent execution environments mean agents that outlive the session. The Ona acquisition signals that “the agent kept working while I was at lunch” is becoming standard, not a novelty. For PMs, that’s relevant for anything long-running - data migrations, multi-step research, report generation against large document sets - where previously you’d have had to babysit a session or break the task into chunks.
On-device translation removes a real coordination barrier. Gemini 3.5 Live Translate’s continuous, low-latency translation - already in production at Grab and Google Meet - is directly applicable to any project with distributed, multilingual stakeholders. Live status calls, supplier negotiations, and site visits with local teams are the obvious candidates.
Frequently Asked Questions
Why was Claude Fable 5 shut down right after launch? The US Commerce Department issued an export-control directive on 12 June, citing a jailbreak that could let the model identify software vulnerabilities - a capability Anthropic says other frontier models already have. Because the order targeted foreign nationals and Anthropic couldn’t selectively restrict its global user base, it disabled the model for all customers worldwide. Other Claude models, including Opus 4.8, were unaffected.
What does the Visa-OpenAI agent payments integration actually let agents do? Agents can execute payments through Visa’s network within parameters the user sets - spend limits, approved merchant categories, and mandatory multi-factor authorisation. It’s not unrestricted spending; it’s a framework for letting an agent transact within boundaries, using tokenised credentials and real-time fraud monitoring.
Should project teams worry about depending on a single AI model or vendor? The Fable 5 suspension is a strong argument for not building critical workflows around a single premium model. Practically, that means designing fallback behaviour (a mid-tier model that can pick up the work if a premium one becomes unavailable) and avoiding contractual or workflow lock-in to one provider for anything where an overnight outage would be disruptive.
The pattern this week is consequence catching up with capability. Models are powerful enough that a government can - and did - pull one from global availability with hours’ notice. Agents are persistent enough that they can keep running unattended, and now capable enough to spend money on your behalf within limits. For project teams, governance - model fallbacks, spend limits, approval chains - now needs to keep pace with what the tools can do.
If an AI agent could spend money on your behalf within limits you set, what’s the first task you’d hand it - and what limit would you put on it?
Yes, AI helped me to write this :)