The Play — Issue #2

Here's what's on my radar.

Two of the most advanced AI models in the world got pulled from the market by government order this month — and one of them came back this week with strings attached. If your AI tools depend on a frontier model API, and most of them do, the last three weeks were a preview of a failure mode nobody put in your contract.

On June 12, the U.S. government ordered Anthropic to suspend access to its two most advanced models, Fable 5 and Mythos 5, for any foreign national anywhere — including Anthropic's own foreign-national employees. The trigger was a report from Amazon researchers describing a technique that bypassed one of Fable 5's cybersecurity safeguards. Anthropic disputed how serious that was, but with 90 minutes to comply and no way to segment access by nationality, it took both models offline for everyone. They stayed dark for 18 days. On June 30, Commerce lifted the controls — after Anthropic agreed to proactively detect and address security risks, work with the government on standards for upcoming models, report malicious activity, and give designated government partners early access to future models with national-security-relevant capabilities. Fable 5 came back online July 1.

Anthropic isn't the only one. OpenAI previewed its new GPT-5.6 lineup on June 26 to about 20 government-precleared partner organizations — not the public, not even its normal API customers — at the administration's request. OpenAI called the arrangement a short-term step and said it doesn't want government-gated releases to become the default. The administration's position, formalized in a recent executive order, is that frontier models with advanced cyber capabilities now warrant formal review before broad release.

Neither Fable 5 nor Mythos 5 nor GPT-5.6 is what's running your firm's AI tools right now. Those are narrowly-gated, most-capable-tier models built for cybersecurity-grade work, not the general-purpose models most legal AI products run on day to day. Nothing in your stack broke this month. But look at how the story ended, not how it started: the price of getting switched back on was a standing government relationship — security cooperation, standards alignment, and early government access baked into future releases. The question AI researchers are now asking out loud: does the government need to approve every frontier model release? That's not a hypothetical anymore. It's the operating environment your vendors build in.

Thomson Reuters' new CoCounsel runs on Anthropic's Claude Agent SDK — the same company's agentic infrastructure, even though the specific restricted models aren't what's powering it today. That's not a reason to worry about CoCounsel specifically. It's a reason to notice that the dependency chain runs through a company the government has now shown it will act against on short notice, and almost nobody's vendor contract has language for what happens if that reaches the model you're actually using.

The Play this week: Map which of your AI tools depend on which underlying model and vendor — don't assume, ask. Then ask each vendor directly whether they have a fallback if their underlying model gets restricted, and get the answer in writing before your next renewal instead of during an outage. If you have foreign-national attorneys or international offices, move on this now — June's directive targeted foreign-national access specifically, so you carry the most direct exposure. Everyone else, this is precautionary. Do it anyway.

Vendors will tell you they're built on "enterprise-grade" infrastructure. That's a claim about uptime and security. It says nothing about what happens when the model underneath gets pulled by a directive nobody outside the company ever sees.

Legora announced this week that its new Agent Pro product is moving to consumption-based pricing — you pay for what the agent actually does, not a flat fee per attorney. A real-time dashboard tracks usage by organization, user, or matter, and admins set their own spending thresholds and alerts. The base Legora Agent stays included for existing customers at no additional cost; it's specifically the more capable Agent Pro tier that's repricing.

The reasoning is sound even if the framing is a little convenient for Legora: a seat license was always a bet that cost-per-user would stay roughly stable, which made sense for software that sat quietly until someone opened it. Agentic tools don't sit quietly. They plan, call models, and keep working until a task is done, and the compute behind that can vary enormously between a quick contract review and a multi-document due diligence pass. Legora isn't the first software company to make this move — Cursor, Clay, and the model providers themselves got there first — but it's the first major legal AI vendor to do it, and Legora has said publicly it expects others to follow.

The Play this week: If you're renewing or signing any AI vendor contract in the next few months, ask directly whether usage-based pricing is coming for that product. Get spend caps, usage dashboards, and audit trails written in now, while you're negotiating, instead of accepting them later as a take-it-or-leave-it change. And shift the internal budget conversation from "how many seats do we need" to "what's our expected usage volume" — it's a different number, and finance is going to ask for it eventually whether you're ready or not.

The vendor that reprices first doesn't just change its own bill. It hands every vendor after it the language they'll use to justify doing the same thing.

That's Issue #2. Go check your vendor contracts. I'll see you next week.