AI Was Fast. Cheap Was the Part That Broke

Most teams think they are buying capability when they buy AI tools.

In practice, they are buying access.

That distinction did not matter much while access felt abundant, pricing felt simple, and the tools were generous enough to disappear into the workflow.

It matters immediately once limits tighten.

This week (May 2026), when GitHub Copilot raised multipliers on Anthropic models by as much as 7x, it looked like a pricing detail.

It was not a pricing detail.

It was a reminder that many teams did not buy cheap intelligence. They bought conditional access.

If you translate those multiplier changes into rough effective monthly capacity, the shift is hard to ignore:

• Sonnet 4.6: 300 messages/month to roughly 33
• Opus 4.6: 100 to roughly 11
• Opus 4.7: 40 to roughly 11

That is roughly an 89% cut in the first two cases and about a 72% cut in the third.

The exact number is less important than what it signals.

A seat that looked generous on paper now buys far less real work.

And once that happens, the conversation changes.

Not because AI stopped being useful.

Because the economics stopped being simple, and the dependency became visible.

The deeper risk is not price, it is dependency

The pricing change matters because it exposes something larger than one expensive month.

Once teams build real workflows around premium AI access, pricing stops being a procurement detail and starts becoming an operating constraint.

Now you are not just buying a tool. You are inheriting:

• vendor policy risk
• throughput uncertainty
• model switching costs
• retraining costs when behavior changes
• budget volatility in the middle of established workflows

That is a more serious shift than “AI got pricier.”

It means a slice of your team’s capability now lives behind rules you do not set.

An engineering lead who budgets premium AI into a sprint now has a new failure mode: the work itself slows down when the useful model becomes constrained halfway through the week.

The market sold cheap intelligence, not just faster autocomplete

A lot of AI adoption was justified with software logic.

Pay a monthly fee. Get more output. Replace part of the labor. Scale it across the team.

That framing worked because the tools felt abundant.

And abundant tools are easy to mistake for permanent infrastructure.

Humans are slower. Humans are expensive. Software scales.

That story was always incomplete.

It assumed three things that were never guaranteed:

• premium access would remain broadly available
• pricing would get smoother over time, not more restrictive
• output speed would translate cleanly into operational efficiency

The model was fast.

The business model was provisional.

The real metric is cost per reliable outcome

Most AI pricing debates still happen at the wrong level.

People compare seat price to salary. They compare a monthly subscription to an hourly rate. They compare raw generation speed to human working speed.

That is not how the work is experienced.

The real unit is not cost per message. It is not even cost per seat.

It is cost per reliable outcome.

That means the full path from prompt to something a business can actually use:

• getting the model enough context
• checking whether the answer is correct
• retrying when it is shallow, brittle, or off-target
• editing for tone, accuracy, or maintainability
• testing the result when the stakes are real
• owning the cleanup when the first pass was fast but messy

This is why AI can feel magical in demos and more expensive in operations.

The output arrives fast. The certainty does not.

In many teams, the human did not disappear. The human moved downstream into review, correction, validation, and accountability.

That still creates leverage.

AI compresses generation time far more reliably than it compresses accountability.

Eleven good turns is not much room for serious work

The Copilot multiplier change matters because it makes effective capacity visible.

Eleven messages may be fine for:

• a quick explanation
• a small refactor
• a rough outline
• one narrow debugging question

It is not much room for work that unfolds through ambiguity.

Serious use often looks more like this:

1. explain the problem
2. realize the model missed context
3. clarify the constraint
4. inspect the answer
5. correct a bad assumption
6. ask for a second approach
7. test the output
8. hit an edge case
9. reframe the task
10. tighten the result
11. patch what still broke

And that is before you count the turns wasted on answers that were plausible but wrong.

The message counts are also fuzzy by design, which is part of the problem.

Once pricing becomes abstract enough that normal users need multiplier math to estimate real capacity, budget planning gets harder. Teams stop buying a simple tool and start managing an opaque resource.

Once capacity becomes opaque, teams stop budgeting software and start budgeting uncertainty.

The price correction is huge, and predictable

The increase feels dramatic because it is dramatic.

It is also not surprising.

Premium model inference is expensive. The strongest models are costly to serve. The market spent the last stretch training customers to expect a lot of capability for a flat monthly number.

That was unlikely to hold forever.

Early in a platform cycle, vendors subsidize behavior. They need adoption. They need habit formation. They need teams to rewire workflows around the product.

Later, the economics get stricter.

Not always because a company is greedy. Often because the original pricing was never the long-term equilibrium.

The mistaken assumption was not that AI worked.

It was that premium AI access would stay underpriced while usage expanded and expectations hardened.

That was never likely to hold.

AI still wins in a lot of places

This is not a case for abandoning AI.

That would be as lazy as the original hype.

AI still has a clear advantage in work that is:

• low-risk
• narrow in scope
• easy to verify
• repetitive
• structurally predictable
• useful even when the first draft is imperfect

That includes things like:

• first-pass summaries
• outline generation
• boilerplate code
• copy variations
• research distillation
• internal drafting support for strong operators

AI does not need to replace humans to be useful.

It only needs to reduce friction where judgment is still retained by someone competent.

The problem starts when teams budget for it as if it removed the need for judgment altogether.

That is where The Illusion of Progress in Modern Web Development becomes relevant. Faster movement is real. The mistake is assuming movement and finished value are the same thing.

The labor comparison comes back in messy workflows

The cleaner the task, the stronger the economics.

The messier the task, the faster the human comparison returns.

AI loses its obvious cost edge when the work is:

• ambiguous
• high-context
• review-heavy
• sensitive to mistakes
• spread across long sessions
• dependent on premium models to stay useful
• expensive to verify after generation

That is where the sentence people do not want to say out loud starts becoming plausible:

In some workflows, it may be cheaper to pay a human than to keep pretending the AI path is still software-cheap.

The point is not that humans are broadly cheaper.

It is that supervision-heavy AI workflows can erase the cost advantage teams assumed would exist by default.

Because the AI workflow now includes hidden labor:

• retries, supervision, testing, and cleanup
• fallback decisions when access or quality shifts

If one experienced person can solve a messy problem in one focused block of work, the economics can beat a premium model that is capped, volatile, and still needs babysitting.

This is especially true when the human owns context the model has to be re-taught every single time.

That is how a pricing change turns into process drag.

Not dramatic failure. Quiet friction. The kind that slows decisions and makes people work around tooling constraints instead of through the problem itself.

Teams already know what this feels like in other forms. It is the same family of invisible cost behind process drift and the same kind of unmanaged tradeoff that often turns into technical debt.

Buy leverage, not the fantasy

The sober takeaway is not anti-AI.

It is anti-fantasy.

AI is still a force multiplier. It is still often worth paying for. It is still changing what one capable person can get done in a day.

What looks weaker now is the lazy budget narrative.

The one that said intelligence had become software-cheap. The one that assumed premium access would stay abundant. The one that treated faster output as proof of better unit economics.

That was the fragile part.

So buy AI for what it actually is:

• variable leverage
• accelerated drafts
• faster exploration
• selective throughput gains
• a useful tool inside a human-owned system

Do not buy it as permanently cheap labor.

The leverage is real.

The illusion was thinking the leverage layer belonged to you.