Renting frontier AI used to be the safe default. This week it became a supplier risk, and open weights became the continuity plan.
At 5:21pm Eastern on a Friday, a one-line Commerce Department letter ordered Anthropic to block every foreign national from Fable 5 and Mythos 5. Within hours, two of the strongest models on the market went dark for every customer outside the US. No warning, no appeal. If you had built a product on Fable 5 from Lagos or Seoul, your app's brain vanished over a long weekend, and the reason was your passport, not your invoice ("Anthropic model ban").
Hold that next to a second fact from the same week, because together they change what renting AI means. Buyers now carry a dependency they have to test before it tests them. Sellers have a continuity product they can build and charge for.
The hedge was already on the shelf. While access to the closed frontier turned into a question of citizenship, the open-weight models kept landing, and they kept winning on the numbers teams actually feel. Moonshot's Kimi K2.7 Code scored 81.1 percent on a tool-use test against 76.4 for Claude Opus 4.8, at roughly a tenth the price ("Kimi K2.7 Code"). Z.ai's GLM-5.2 beat GPT-5.5 on real coding benchmarks at a sixth the cost, weights posted to Hugging Face under a permissive license ("GLM-5.2"). DeepSeek-V4-Pro now tops the open SWE-bench leaderboard at around 0.81, priced near 87 cents per million output tokens ("DeepSeek-V4-Pro"). MiniMax shipped M3 with a million-token context window, and one of this week's issues named the moment plainly: a self-host hedge against the ban ("MiniMax M3").
Treat those scoreboards with the caution they earn. MiniMax's 59 percent on SWE-Bench Pro is a vendor-run figure that skips the newer Opus 4.8. Snowflake says its CoCo agent beat Claude Code and OpenAI Codex on a data-engineering test, on its own numbers, which nobody outside the company has reproduced ("Snowflake CoCo"). The point is not that the open models have won. It is that the gap closed far enough that price and control, not raw capability, now decide what many teams build on. That is a different world from the one where you paid the frontier premium because nothing else came close.
The money is moving to match. Prem, a Swiss startup that helps banks and law firms run models on their own infrastructure, is raising $100 million the same week the Anthropic block reminded everyone that rented access can disappear ("Private AI: Prem"). Sovereign and on-prem AI stopped being a compliance footnote and became a funded thesis.
Here is the shift underneath all of it. On a hosted model, you do not own anything. You rent a permission, and a permission can be revoked, repriced, or retired on a calendar you do not control.
This week supplied the whole catalogue. OpenAI pulled GPT-5.2 and its Codex variant, forcing a migration to 5.5 ("OpenAI deprecation"). Google retired the Gemini CLI and pushed developers to Antigravity, with several image and video API endpoints scheduled to shut off on June 25 and 30 ("Gemini CLI retired," "Gemini API cutoffs"). GitHub's Copilot moved from flat per-request pricing to metered, token-based billing, so the cost of an agent reasoning over a large repository now swings with every long session ("Copilot's metered billing"). Anthropic paused its plan to meter Agent SDK usage, but paused is not cancelled ("Claude Agent SDK billing"). Each of these is a vendor changing the terms of a dependency you already shipped.
The independent lane closed at the same time. SpaceX bought Cursor's maker for $60 billion in stock, which means every serious AI coding assistant now sits inside a tech giant: Copilot at Microsoft, Claude Code at Anthropic, Codex at OpenAI, Cursor at SpaceX ("Cursor sold"). The neutral tool you standardized on is now somebody's strategic asset, and its pricing and roadmap answer to that owner.
And the prices you rent at may not be the real prices. OpenAI's leaked books showed a $38.5 billion loss against $13.1 billion in revenue, most of it a one-time charge, with an operating loss still near $8 billion ("OpenAI's books leak"). The frontier rate card is held up by investor money while these companies file toward public listings. When the subsidy tightens, the bill you modeled on this year's pricing is the optimistic case.
So retire one assumption. The old default said the best AI lives behind the frontier API, and renting it is the safe, boring choice. This week made the boring choice the risky one. The safer move is to treat every frontier model call as a supplier relationship with a second source on standby, the way you would never run a factory on a single-source part you could not substitute.
For buyers and operators, run a fallback drill next week. Take one workflow that leans on a frontier model, stand up GLM-5.2 or DeepSeek-V4-Pro on rented GPUs, run your real evals on both, and compare cost per successful task, error rate, and where each one breaks. You are not switching vendors. You are buying the right to switch, and pricing what it would cost before someone else picks the date.
For sellers, consultants, and software teams, there is a clean offer in this. Package a model-dependency audit: map every frontier call in a client's stack, score each one for revocation risk across the vendor's deprecation history, the client's jurisdiction, and the billing model, wire a single fallback route to a self-hostable open weight, and hand over the eval suite that proves the fallback holds on the client's own tasks. The Anthropic block just handed every regulated buyer the business case to pay for it.
The companies selling intelligence would like you to treat access like electricity, a settled utility that is simply always on. This week showed it is closer to a lease you can be evicted from. Keep one model you can run yourself, even if you never have to.
The week in one line: Treat every frontier model as a rented permission with a revocation date you do not set, and keep one self-hostable open weight ready as the continuity plan, not a downgrade.
Sources this week: Anthropic model ban, Kimi K2.7 Code, GLM-5.2, DeepSeek-V4-Pro, MiniMax M3, Prem sovereign raise, OpenAI deprecation, Gemini CLI retired, Copilot metered billing, Claude Agent SDK billing, Cursor sold, OpenAI books leak