You are being set up.
Not in a shadowy-cabal way. In the very boring, venture capital way: subsidize the product, let people wire it into their daily work, then extract big profits later by jacking up prices.
That trap product is AI. The cheap, all-you-can-eat AI you are building on right now is bait.
Enjoy it. Use it. Ship with it. I certainly am. But do not build your dev process on the assumption that frontier-model access will stay cheap, generous, and casually available forever. That assumption is a rug, and I think it is going to get pulled soon.
If you are an engineer or engineering leader, this is not abstract macroeconomics; it lands in your agent runs, your CI loops, and eventually your delivery margin. If your workflow only works when the agent can spray tokens at the most powerful model all day, with no capacity friction, you are not adopting AI. You are accepting a subsidy.
That can be fine. Subsidies are useful. Just don't mistake it for infrastructure.
Name the play
The playbook is called penetration pricing. Sell below your long-term profit target until the market forms around you. Get the workflows, the defaults, the "well, everyone already uses this" muscle memory. Then, once switching is painful, jack up the price.
Plenty of useful products started this way. The early pricing is customer acquisition, not a promise about the future.
The weird part is what gets subsidized: not a tidy SaaS app with predictable costs, but something compute-hungry, habit-forming, and woven into how technical teams think, plan, code, and review. So when the terms change, it will not feel like your streaming bill went up. It will feel like your development process got a new tax.
The clock is counting down
Anthropic confidentially filed its S-1 on June 1, 2026, at a reported valuation around $965 billion. OpenAI reportedly filed about a week later, around June 8, 2026, with analysts eyeing a late-2026 or early-2027 listing. Two major AI providers heading toward public markets within days of each other is not a coincidence. It is a timer.
Public markets are not famously patient about profits; they want to know when the thing prints cash instead of burning it.
OpenAI's own forecast reportedly points to a roughly $14 billion loss in 2026. Anthropic burns less, but it still burns: it reportedly expects to lose around $3 billion in 2026 (down from about $5.6 billion in 2025), and to stop burning cash only in 2027. Both are growing fast, into the tens of billions in run-rate, and both carry near-trillion-dollar valuations that assume the margin eventually shows up. You can grow fast and still be bleeding margin.
When a provider needs margin, the levers are obvious: raise prices, gate the best models, ration heavy users. Which appears first? Rationing, and that's already happening.
Anthropic introduced weekly rate limits on Claude Pro and Max in late August 2025, across the $20, $100, and $200 tiers, saying they would hit less than 5% of subscribers.
The first to hit the wall are the ones showing what heavy usage really costs: developers and agent-heavy teams. You find where the expensive users live, then build fences around that behavior.
It's happened before
None of this is mysterious, because the pattern is old.
Uber went public in 2019 at about an $82 billion valuation while losing more than $3 billion a year, then made the cheap, frictionless default less cheap once the market was formed. Netflix marched its standard plan from around $7.99 before 2019 up through staged hikes to roughly $19.99 by 2026, with premium near $26.99. Capture the habit, become the default, then raise the price.
The new wrinkle is how dependent orgs are becoming. You can quit a streaming service or go back to taxis. But if your delivery process assumes abundant frontier-model capacity, a big price hike stops being an annoyance and starts dictating what you can ship.
Fable is the bait in its purest form
The cleanest example, to me, is Fable.
Fable is the consumer version of Mythos, a model tier above Opus: the shiniest, most compute-hungry class available. It showed up for only a few days recently, then got pulled. While it was live it ran at twice the price of Opus (about $10 per million tokens against Opus's $5), and was almost certainly under-priced even then. The top tier changes what you are willing to ask the machine to do.
You try the expensive model. It handles the ambiguous ticket better, holds more of the plan in its head, and untangles the gnarly legacy cartridge code that makes smaller models stall.
And then your brain does the dangerous thing. It updates the baseline.
Suddenly the workflow is not "use AI thoughtfully." It is "route everything through the biggest model I can reach." You stop asking whether the task needed the top tier, or how much context you dumped in, because the model swallowed it yesterday and nobody screamed.
That is the trap: the most dazzling model is the easiest to build bad habits around. If Fable, or anything like it, comes back as a premium, scarce, tightly metered capability, the teams most excited by it will be the teams most exposed to it.
The move is not "use less AI"
I am not saying retreat to hand-coding everything because AI economics got scary. The move is to refuse capture.
Use AI aggressively while access is cheap, but build as if the best model might get scarce or expensive next quarter, because the rug pull is coming soon. Once these companies IPO, watch out.
In practice: write specs tight enough that the agent need not infer the universe from a one-line Jira ticket and three angry comments; keep context lean; use the lowest powered model you need; and avoid vendor lock-in no matter what.
I also wrote a companion piece on cutting AI coding-agent token costs; that post is the practical "how," this one is the "why you should care before finance cares for you."
And yes, this is where Bridge GPT fits: spec-driven SFCC delivery that shapes tickets, without locking you into a vendor, so you can pivot where you need.
The teams that prepared will shrug and reroute. The teams that took the bait will discover how much of their process was rented.
Build accordingly.