The Strongest AI Models Now Need an Invitation. Your Software Vendor Will Get One Before You Do.

If your marketing stack runs on AI, the model doing the work is increasingly not a thing you choose. Your software vendor chooses it, and the vendor’s own access is now something it has to be granted.

That’s the shift Tomasz Tunguz of Theory Ventures described on August 31, and the reason to care isn’t philosophical. It changes what “switching tools” means, and it puts a dependency in your stack that no contract you signed mentions.

His line:

“Access, not price, is the new scarcity at the frontier.”

A frontier model is just the current strongest model a lab has built, the one that can do work the cheaper ones can’t. For two years those arrived as a utility: pay per use, anyone with a credit card gets the same thing. Tunguz argues that era is closing.

The proof is on Anthropic’s own page

You don’t have to take a venture capitalist’s word for it. Anthropic’s Claude Mythos page says it plainly:

“Claude Mythos 5 is currently only available to a small group of vetted partners with a goal of opening up more broadly in the future.”

And: “We currently only provide access to a small, but growing, set of customers through our trusted access programs.”

The timeline on that page is the part worth writing down. Mythos 5 was announced June 9, 2026. It became unavailable on June 12, three days later. On July 1, export controls were lifted and access was restored, in the page’s own framing, for a set of US organizations.

A model that appeared, disappeared inside a week for reasons no customer controlled, then came back for an approved list. Using it also requires accepting a 30-day data retention policy for safety monitoring, which is its own consideration if you’re feeding it client material.

Tunguz reads the same pattern into the wider market: “Rationing programs & export limits decide who may run the strongest models,” and “Labs are developing whitelists & blacklists.”

Where it reaches your stack

The second half of his argument is the half that lands on a marketing team:

“Downstream, enterprises standardize on one or two named vendors, & products ship with a default model inside.”

Salesforce made that concrete five days before he wrote it. Its Claudeforce announcement on August 26 puts Claude inside Agentforce Vibes, Agentforce Coworker and Slackbot by default, and makes it a reasoning model for the Atlas Reasoning Engine. Marc Benioff framed it as a merger of category leaders: “We’re bringing together the world’s #1 AI and #1 CRM — the best of both worlds.”

Read the announcement for what it doesn’t say. It never addresses whether a customer can run a different model in those surfaces. Tunguz is blunter about what that pattern means: “SaaS apps pick their components & pass them on a single app to the customer : take-it-or-leave-it.”

Nobody buying Salesforce is being mistreated here. Claude is a strong model and a default is a reasonable product decision. The point is narrower: a choice that used to be yours moved one layer up, and you’ll find out it moved when it changes.

Why this is different from normal vendor lock-in

Switching costs are old news. Two things here aren’t.

The supply can be revoked by someone who isn’t your vendor. Mythos 5 went dark on export controls, not on a business decision. We covered the commercial version of this last week when OpenAI ended Cursor’s model access over who bought Cursor, where the trigger was an ownership change the customer had no part in. Government action and a change-of-control clause are different mechanisms with the same result: the model in your tool can leave for reasons that have nothing to do with you.

The dependency is invisible in procurement. Nothing in a martech contract typically names which model powers a feature, which lab supplies it, or what happens if that supply stops. You can’t audit a dependency you were never told about.

We wrote in early August about rationing arriving on the cost side, when Uber burned its annual AI budget in four months. Access rationing is the same squeeze from the other direction. One limits what you can afford to run, the other limits what you’re permitted to run.

Three questions worth asking this quarter

None of this calls for panic, and switching your stack over it would be an overreaction. It does call for knowing what you’ve got.

Which models power the AI features you already pay for? Most teams can’t answer this for a single tool. Ask your vendors in writing. The answer is also a reasonable thing to want in a renewal.

What happens if that model becomes unavailable? Ask whether the vendor has a fallback and whether switching would change output quality. A vendor with a real answer has thought about it. A vendor without one is telling you something too.

Where does model choice matter to you? For most marketing work it doesn’t; a default is fine and the productivity is real. Find the one or two places where output quality is tied to revenue, and keep a route that isn’t controlled by a single supplier.

The useful takeaway is smaller than the headline. Two years of treating AI capability as a utility trained everyone to assume the strongest model is a purchase decision. On Anthropic’s own page today it’s an eligibility decision, and the entity being judged eligible is usually your vendor rather than you.