Google tripled its cheapest tier, Anthropic locked its best model, and even Apple rents its brain: in AI the vendor holds the option.
Apple builds its own chips. It designs the silicon in your iPhone, your laptop, and the watch on your wrist. This week, at Tim Cook's final WWDC keynote, Apple confirmed the new Siri runs on a 1.2-trillion-parameter model Google built for it, under a licensing deal worth about a billion dollars a year. The company most famous for owning its whole stack is now renting the part that matters most.
That detail, "Apple WWDC: Siri rebuilt on Google's Gemini," is easy to file as a one-off. It isn't. Read against the rest of the week, it is the clearest sign of where leverage actually sits, and it is not with the buyer.
Look at what the platform owners did with that leverage in the same seven days. Google retired Gemini 2.0 Flash and its lite variant on June 1, and the only migration path, Gemini 3.5 Flash, costs three times the tier it replaced. If you built a product's margins on the old price floor, your API bill tripled without you touching a line of code. Google's defense is that the new model runs twelve times faster on its own benchmarks. That can be true and still beside the point, because the number that changed for you was the invoice, not the latency.
The obvious objection is that prices are falling, not rising. DeepSeek raised $7.4 billion in its first outside round, Moonshot is chasing a $30 billion valuation, and that wave of cheap, capable Chinese models is real downward pressure on what US labs can charge. But a falling average and a stable price for the specific tier you depend on are not the same thing. The floor can keep dropping while the rung you are standing on gets pulled.
Then the gating. Anthropic shipped two models on June 9. Developers got Claude Fable 5. Its stronger sibling, "Claude Mythos 5," went out locked behind a clearance review, available only to vetted critical-infrastructure and biology partners. OpenAI did the same with biology: GPT-Rosalind got cheaper and better at genomics, then shipped with a separate biodefense tier handed only to approved agencies. The pattern is plain. The best capability is no longer something you buy. It is something you are granted.
The inversion under all of it. The thing every buyer assumed about this market, that more vendors and cheaper tokens mean more power in your hands, is being quietly turned over. In AI, the vendor keeps the option. You hold the obligation.
The buyer the seller can punish. Watch the Pentagon. The Defense Department is now testing OpenAI, Google, and Grok to replace Claude on classified systems, after Anthropic refused to drop guardrails on mass surveillance and autonomous weapons and got labeled a supply-chain risk for it. Read that next to Microsoft. At Build, Microsoft launched seven in-house MAI models and claimed one beat OpenAI's GPT-5.5 at a tenth the cost in a McKinsey deployment. Take the tenfold figure with salt; it is one task, one customer, the vendor's own keynote. The strategic message under the benchmark is the real news: Microsoft now owns the silicon, the models, and the apps, and no longer has to recommend anyone else. The same week, it pushed agent governance down into the operating system with the Agent 365 SDK, so the controls that decide what your agent may touch live in a platform you rent, not code you keep. Who holds the off switch is the question, and the answer is drifting away from you.
Who lost the option to wait. Walk this down to the actual work. OpenAI shipped Codex for every role, turning the analyst's and the junior banker's workflow into a plugin; non-developers are already a fifth of its five million weekly users and growing three times faster than the engineers it was built for. AlphaSense doubled to a $7.5 billion valuation by selling the exact task a first-year equity research associate used to do, reading every filing to find the one number that moves a thesis. And AI was named in 38,579 US job cuts in May, forty percent of the month's total. Whether the software truly caused those cuts or just gave a tidy label to decisions already made, the entry tier is thinning either way. It is the same machine seen from a different floor: the platform owner sets your infrastructure price and absorbs the junior headcount you used to train into senior judgment.
There is a tell in how the labs talk about this. The same week it shipped a gated frontier model, Anthropic's policy team warned that self-improving AI is close and asked the industry to build a brake pedal, noting Claude already writes 80% of Anthropic's own code. Sit with the shape of that. The company building one of the fastest models is asking everyone to agree, together, where the brake should go. The candor is real, and so is the position it protects.
So here is the assumption to retire. For a decade, buying software meant a contract that held: a price, a version, terms that changed on notice you could plan around. The reflex carried into AI, where teams treat a model the way they treat a database, a dependency that sits still while they build on top of it. This week says it does not. A model is a leased asset with a notice period the vendor controls, and the lease can be repriced, retired, or revoked between a Monday and a Friday.
Two things follow. Stop pricing your product on the cheapest tier available today as if it will be there next quarter; price as if it will be tripled or pulled, and keep a tested second model you can fail over to, so a forced migration is a configuration change instead of a fire. And write your own evaluations against your real workload, the way Google's move to run Kaggle Benchmarks locally now lets you do from the editor you already use, so when a vendor reprices or gates its best model you can tell in an afternoon whether the alternative holds, rather than taking the next keynote's word for it.
The leverage in this market was never going to sit with whoever shipped the cleverest model. It sits with whoever can walk away. Apple couldn't, and it is paying a billion a year. The work worth doing now is making sure you can.
The week in one line: Treat every model as a leased asset the vendor can reprice or revoke on its own clock: price as if today's cheapest tier disappears, and keep a tested fallback plus your own evals.
Sources this week: Apple's Siri on Gemini, Gemini 2.0 Flash retired, DeepSeek's first round, Moonshot at $30B, Claude Mythos 5 gated, GPT-Rosalind, Pentagon tests Claude rivals, Microsoft MAI models, Microsoft Agent 365 SDK, Codex for every role, AlphaSense at $7.5B, AI tops US job cuts, Anthropic's brake pedal, Kaggle Benchmarks local