Edition #14. This week the IMF came out with an unusually firm warning: new AI models like Claude Mythos can cause a 'systemic shock' to the financial system, according to the fund. At the same time, Anthropic struck a deal with SpaceX for 220,000 NVIDIA GPUs within a month, which immediately doubled the 5h-window in Claude Code. OpenAI shipped three new realtime voice models with GPT-5-level reasoning and 128,000 tokens of context per conversation. And in fintech it became visible what AI does to work: Coinbase laid off 700 people, or 14 percent of staff. Four stories that this week hit the regulators, the compute market and the labor market all at once. Let's dive in!
The IMF publicly raises the alarm about a specific AI model for the first time
On Thursday the IMF published a blog post in which senior officials warn that the latest AI models "elevate cyber risk to a potential macro-financial shock". The trigger: Claude Mythos from Anthropic, which in its controlled release has found thousands of high-severity vulnerabilities in virtually every major operating system and web browser.
Anthropic has rolled Mythos out to forty mostly American organizations, including Amazon, Microsoft and JPMorgan Chase. Non-American banks aren't part of those forty and don't have access to the same protection. "Cyber risk does not respect borders," the fund writes, with explicit concern for emerging markets.
What makes this unusual: the IMF for the first time names a specific AI model in a warning about financial stability. Until now, AI cybersecurity was an abstract talk for central bankers. Mythos has changed that. The fund officials close with a line that lingers: defenses will inevitably be breached, so resilience matters more than prevention. For anyone thinking about AI strategy, that's a clear signal.
Anthropic rents 300 megawatts of compute from SpaceX, and you feel it directly in Claude Code
On Tuesday Anthropic announced a partnership with SpaceX to use the full compute capacity of the Colossus 1 data center: 300 megawatts and around 220,000 NVIDIA GPUs, available within a month. Concretely for Claude Code users: the 5h-window has doubled (and the extra reduction during peak hours is gone). On top of that, the API rate limits for Opus are substantially up.
This is not a one-off deal. Anthropic has also closed capacity deals with Amazon, Google, Microsoft/NVIDIA and a $50 billion infrastructure investment with Fluidstack in recent months. SpaceX was until now mostly a customer and co-owner of xAI's compute. That Anthropic now gets to use that capacity says something about how scarce compute has become.
The practical impact is immediate if you use Claude Code daily: the limits regularly got in the way during big refactors or long agent runs, and that's been all but solved since Tuesday. For teams that have Claude in their workflows, this is a free upgrade.
OpenAI brings voice agents to GPT-5 level, with live translation and streaming transcription
On the same day, OpenAI shipped three new realtime audio models in its Realtime API. GPT-Realtime-2 is the headliner: GPT-5-level reasoning, can call tools, recovers cleanly from interruptions, and runs with a 128,000-token context, enough for an hour-long customer call with full history. They also launched GPT-Realtime-Translate (live translation from 70+ input languages to 13 output languages) and GPT-Realtime-Whisper (streaming transcription as people speak).
The pattern OpenAI itself names: listen, reason, translate, transcribe, act. For the first time, a complete voice-agent stack sits in a single API. Until now, you had to wire three to five services together for a serious voicebot, with latency that broke the conversation. That barrier is gone with this release.
What I think this does is open a second round of AI in call centers. The first round, with IVR bots and simple scripts, never really worked: too dumb, too frustrating for customers. With reasoning at this level, it gets serious. For customer service and telephony, this is now a real alternative.
Coinbase, Block and Kraken: AI now does the work compliance teams used to do
Coinbase laid off 700 people this week, 14 percent of staff. Block and Kraken had their own rounds, with hundreds of jobs per company. The market climate is part of it: trading volumes sit below the peaks of 2024. But Coinbase's official communication explicitly cites AI: "improving operational speed while aligning with increased use of artificial intelligence in product and engineering work".
What stands out: the work that disappears isn't in the typically vulnerable jobs. It's compliance, transaction screening and customer support. Tasks where regulation, accuracy and high volumes meet, and where AI has broken through over the past 18 months. This isn't Klarna-style "we replaced support with GPT-4", but a structural move toward smaller teams and automated workflows.
Fintech is almost always the first sector where you see what's about to happen to the rest of office work. Insurers, banks and accountants do similar work. What I tell companies thinking about AI: make sure your people are better with AI than they are without it, otherwise a year from now you'll be on the other side of one of these press releases.
What I built this week: calling my OpenClaw agent
This week I hooked up GPT-Realtime-2 via Twilio to my OpenClaw AI agent. Result: I can now call my agent on a phone number and it picks up, with the same reasoning and tool use as the text version. For me this is the first moment a voicebot doesn't feel like an IVR. The agent can look up a database mid-conversation, schedule an appointment and call the customer back. Twilio plus the Realtime API is literally a few hundred lines of code. The kind of custom agent we needed months for not long ago, you now build in a week.
Want to work smarter with AI?
Want to work smarter with AI, in your business or just for yourself?
At AI Agent B.V. we help companies and individuals with AI advisory, AI implementation and practical training. So not just nice stories about what AI can eventually do, but concretely helping with how to use AI smarter today in your work, processes and daily operation. From strategy and tooling choices to training, implementation and custom AI agents.
Interested? Schedule a free 30-minute call.
