The Ladder Pulled Up: AI Subscriptions After the Subsidy

In April 2026, the frontier AI labs did not run out of capability. They ran into economics. What changed was not the intelligence on offer, but the price of asking for it at workflow scale.

Listen to this article (11 min)

0:00

/671

1×

Dublin Airport to City Centre, Again

For a stretch in the mid-2010s, an Uber from Dublin Airport into town could feel less like a fare than a temporary argument with reality.

Not "good value" cheap. Suspiciously cheap. The ride was real. The convenience was real. The price was the fiction.

The interesting question was never whether that phase would end. Venture subsidies always end. The interesting question was what happens after a large group of people has reorganised its habits around a price that was never going to hold.

I have been thinking about that question again this month, at a different dashboard, for a different kind of fare.

In the first week of April, AI subscriptions started to feel the same way. Not broken. Not unavailable. Newly honest.

Anthropic restricted subscription access through third-party agentic tools, and OpenAI further separated casual use from sustained coding use through new tiers, credits, and Codex pricing. Read together, those changes say something precise: frontier AI is still cheap for ordinary interactive use, but it is getting steadily less cheap for long-running, workflow-level cognition.

That distinction matters. If you use these tools like chat, the old subscription logic still mostly holds. If you use them like infrastructure, the meter is becoming visible.

What Actually Changed

Anthropic signalled the direction before it made the harder cut. In its own usage-limits guidance, the company began explicitly steering heavy users away from weekday peak windows, defined as 5 AM to 11 AM Pacific, and offered roughly double usage outside those hours. That is not a small UX tweak. It is a capacity signal: sustained coding workloads are expensive enough that time of day now materially affects what a subscription buys.

Then in early April, Anthropic made the policy distinction explicit. In its own help-centre guidance on subscription authentication, the company says subscription plans are designed for ordinary use of native Anthropic applications and that third-party tools should use API key authentication through Claude Console or a supported cloud provider. TechCrunch reported on 4 April that Anthropic had begun enforcing that distinction against external agent harnesses such as OpenClaw.

OpenAI moved through product design rather than prohibition. On 2 April, it introduced flexible Codex pricing for ChatGPT teams, pushing heavier coding usage toward seat-based and usage-based billing. On 9 April, OpenAI's ChatGPT release notes added a new $100 Pro tier for longer Codex sessions and temporarily expanded Pro usage relative to Plus. In parallel, OpenAI made flexible usage credits and top-ups a normal part of the product for users who exceed plan limits.

Anthropic's message was blunt: this class of usage no longer fits inside the subscription. OpenAI's message was smoother but economically similar: this class of usage now has a clearer price.

None of this is dramatic in isolation. Together, it amounts to the same move executed in different styles. Anthropic made heavy subscription use more conditional in time and narrower in scope. OpenAI created clearer paid lanes for sustained coding work. The common point is not that the companies behaved identically. It is that both are drawing a sharper line between "person using an assistant" and "workflow consuming compute."

What the Platforms Are Optimising For

You do not need leaked margins or dramatic loss estimates to understand why this line is being drawn. The product logic is enough.

A flat subscription works beautifully for the median user: some prompting, some browsing, occasional coding help, maybe a longer session now and then. That user is predictable. Their usage spikes, but it does not run continuously. A subscription model can absorb them comfortably.

The agentic user is different. Give a strong model a filesystem, tool access, a large context window, and permission to iterate, and you are no longer selling "chat." You are selling sustained inference, orchestration, and repeated tool calls wrapped in the language of a chat plan. That is much closer to infrastructure consumption than to software-as-a-service in its old form.

The difference is not philosophical. It is operational. A casual user asks a model a question. A heavy user asks it to stay awake: read fifteen files, run the tests, revise the patch, retry on failure, explain the diff, then keep going. The moment a subscription starts caring what hour of the day you do that, you are not really buying software in the old sense. You are buying access to constrained compute.

This is why the April changes matter. They do not mean the cheap era is over for everyone. They mean the labs are narrowing the category of use cases that can still hide inside a flat monthly price.

That is a more defensible and more important claim than "AI got expensive." It did not, uniformly. What got more expensive is the habit of treating a consumer subscription as if it were a low-friction compute budget for serious automation.

Europe Pays in Different Hours

There is a detail in Anthropic's peak-hours model that is immediately obvious from Dublin and much less obvious from San Francisco.

Anthropic's weekday peak window, 5 AM to 11 AM Pacific, maps to 1 PM to 7 PM in Dublin in April. In other words: the most constrained part of the subscription day lands directly on the European afternoon, the hours when a data engineer is usually doing the hardest work.

That asymmetry is not malicious. It follows geography, data-centre concentration, and where the largest customer base currently sits. But it is still an asymmetry. Engineers in Europe can pay the same sticker price and receive a product whose tightest constraints arrive in the middle of their working day.

For Irish teams, there is an extra twist. We live in one of the countries most entangled with the infrastructure and legal footprint of US tech, yet the practical shape of these products is still set elsewhere. When subscription plans get tighter, the US working day is treated as normal demand and the European afternoon becomes a congestion problem.

That is not a moral indictment. It is an operational fact. If you are in Dublin, "peak hours" is not an abstraction. Peak hours is after lunch.

What the Cheap Era Built

It is easy to make this a story about pricing and miss the more interesting point, which is what the underpriced phase enabled.

For roughly two years, individuals and very small teams had access to frontier cognitive capability at a price point that behaved more like a streaming subscription than a professional services budget. That mattered.

It meant engineers could learn agentic patterns by doing rather than by reading vendor material. It meant freelancers could prototype above their weight class. It meant small internal teams could test workflows before asking procurement for permission. It meant people transitioning into software and data work could spend long hours with real tools instead of toy demos.

It also meant, in my case, that a small publishing operation in Dublin could build an agentic content pipeline at consumer-plan economics. This site is partly a product of that window.

The capability is not disappearing. The models are stronger than they were a year ago. The change is in who can afford to use them wastefully, experimentally, or continuously without first turning the question into a budget line.

That last part matters because experimentation is not evenly distributed. The large organisation can absorb a metered bill more easily than the individual learner can absorb uncertainty.

What Changes for Practitioners

If you run engineering or data workflows against these systems, the April reset has a few practical consequences.

First, separate interactive use from batch or agentic use in your own mental model. If a workflow can run for tens of minutes, touch many files, or iterate through repeated tool calls, stop treating subscription access as the baseline assumption. Model a direct API path and price it before the platform forces the migration.

Second, add observability around AI-assisted work. Track cost per run, wall-clock duration, file-touch count, and retry loops. Most teams still talk about AI usage in anecdotes. That is not enough once credits, overages, and throttling enter the picture.

Third, if your team is in Europe and still depends on subscription capacity, schedule deliberately. Use interactive daytime access for genuinely interactive tasks. Push longer coding or review jobs into asynchronous runs before 1 PM or after 7 PM Dublin time when the platform's own rules make that materially cheaper in practice.

Fourth, treat third-party agent harnesses as provisional unless they have a clean API-backed operating model. The fragile architecture is not "agentic AI." The fragile architecture is building a serious workflow on top of a subscription entitlement that the provider never intended to underwrite.

Fifth, tighten prompt and context discipline. Route lighter tasks to cheaper models. Keep contexts narrow. Cache reusable work. In the past year those were signs of taste. Now they are basic cost controls.

The broader adjustment is cultural: teams need to stop confusing convenience pricing with durable economics.

The Ladder, More Precisely

The companies making these changes are not behaving irrationally or unusually. They are behaving like platforms reaching for a cleaner fit between price and usage.

What is worth noticing is who gets squeezed first.

When a subsidised resource is repriced toward its true operating cost, large organisations usually adapt by moving the spend into procurement. Individual users and small teams adapt by using less, experimenting less freely, or not crossing the threshold at all. The redistribution is predictable even when nobody intends it.

That is why the ladder metaphor still feels right to me, but only if stated carefully. The ladder is not gone. The labs have not withdrawn capability from the world. What they are withdrawing is the unusually generous mismatch between what serious users could consume and what they had to pay.

If you spent the last two years using that mismatch to build skills, ship projects, or establish leverage, good. That was the correct move. If you are only now arriving, the door is still open, but it is becoming a door that expects a budget behind it.

This is the part the industry often describes too politely. Repricing does not just change margins. It changes who gets to learn by doing.

The immediate response is not nostalgia or despair. It is precision. Know which of your workflows justify metered spend. Know which still fit inside subscriptions. Know when you are consuming an assistant and when you are quietly renting infrastructure.

For a brief, strange period, frontier cognition was sold with the emotional logic of a streaming subscription and the physical cost structure of a compute cluster. That was never going to last. April 2026 did not end the capability. It priced the mismatch.

Simon Cullen is a data engineer based in Dublin. He writes about AI infrastructure, data systems, and the intersection of technology and professional practice at insights.simon-cullen.com.