AI's price collapses as the race shifts from smartest to cheapest

Information Technology 4 min 80 sources

Anthropic's Opus 5 lands at half the cost, Microsoft's own models undercut OpenAI by up to 89%, and Google ships a cheaper Gemini - the frontier's pricing power is eroding fast. Meanwhile AI startups now out-lobby defense giants, custom chips chase Nvidia, and Brussels fines Google $1 billion.

Key takeaways

The AI race is turning from who has the smartest model to who has the cheapest good-enough one, and each cheaper release erodes the pricing power of the expensive frontier - even the vendor's own last model.
The shift shows up everywhere at once: Anthropic, Microsoft, and Google all cutting prices, AI startups now out-lobbying defense giants, and custom chips circling Nvidia.
Brussels fined Google about $1 billion under Europe's new competition law, and the ruling immediately became a US-EU trade fight.

What happened

The most expensive thing in tech got cheaper this week, on purpose, from every direction at once.

On Friday, Anthropic launched Claude Opus 5 [9]. The company says it delivers nearly all the intelligence of its top model, Fable 5, at half the cost [9]. Its own framing is the tell: the AI race is “shifting from raw capability to the economics of daily use” [9]. Opus 5 arrived just two months after the model it replaces [79]. Businesses are watching their AI bills, and the vendors know it [7].

Two days earlier, Microsoft made the same bet from a different seat. It shipped two in-house models and published production numbers claiming they cut costs up to 89% versus OpenAI [3]. These are not lab demos. Microsoft says they now run inside Bing, PowerPoint, Excel, OneDrive, GitHub Copilot, and Azure [3]. The stated goal: “Microsoft products, powered by Microsoft models” [3]. That is a message aimed at OpenAI, the partner Microsoft has spent billions backing [3].

Google fills the third corner. Last week it revealed Gemini 3.6 Flash, pitched as faster and cheaper [14], while its long-promised premium model, Gemini 3.5 Pro, stayed stuck in testing [24].

Why this is happening

For three years the race was about who had the smartest model. That contest is quietly ending, because the smartest model is no longer the one most people need to pay for.

The reason is a claim worth stating plainly: for most real work, near-frontier intelligence delivered cheaply beats frontier intelligence delivered expensively [9]. Anthropic priced Opus 5 at $5 per million input tokens and $25 per million output tokens [9] - a “token” is the unit AI is billed in, roughly a word-fragment. Once a merely-very-good model costs half as much as the best one, the best one loses its grip on the daily workload. The frontier still matters for the hardest jobs; it just stops being where the money is.

The influence economy tilts

You can watch the same shift in an unlikely place: Washington lobbying disclosures.

OpenAI and Anthropic together spent $3.17 million lobbying the US federal government last quarter, a record for both and up 23% from the quarter before [67]. Anthropic alone spent $1.97 million - more than Nvidia [67]. Over the same months, the big established tech firms held their spending roughly flat, and eight major defense contractors cut their combined lobbying by more than 3% [67].

Lobbying spend is a rough gauge of who thinks the ground is shifting their way. Three years ago these AI firms barely registered in Washington. Now two of them out-spend a chip giant and out-climb the defense industry that has owned that corridor for decades [67].

The chip layer, too

The same demolition is reaching hardware. Etched, an AI-chip startup founded by three Harvard dropouts in 2022, raised $300 million at a $10.3 billion valuation - doubling its worth in seven months [1]. Its pitch is a chip built for one job, running today’s dominant AI design, rather than the flexible general-purpose graphics chips Nvidia sells [1]. Backers include Sequoia, Andreessen Horowitz, and SK Hynix [1].

Etched is not alone in aiming at Nvidia. AMD unveiled a rack-scale AI system called Helios [62], and Google is reportedly building its own chip to run models more efficiently [63]. None has dethroned anyone. But a healthy, dominant incumbent is now being circled from several directions at once.

Brussels fines Google $1 billion

The week’s biggest non-AI story: EU regulators fined Google roughly $1 billion (890 million euros) for competition breaches tied to its search and ad businesses [35][40]. It is an early real test of Europe’s landmark digital-competition law [35]. Google plans to challenge it [77].

The political fallout landed fast. President Trump called the fine “illegal,” threatened a “substantial tariff” on the EU for “robbing” US tech firms, and said the US would open a trade investigation in response [71][49][80]. A single antitrust ruling in Brussels is now a transatlantic trade dispute.

Quick hits

Two big breaches. The AI music service Suno was breached, exposing data on about 55 million users [31]. Separately, Hugging Face - a hub where developers share AI models - confirmed a breach hit internal datasets and credentials, and urged users to rotate their keys [42].
DeepSeek hits pause. The Chinese AI lab DeepSeek quietly suspended its second fundraising round, days after viral posts attributed to its founder [15]. Its first round closed in June at $7 billion [15]. The lab is also said to be preparing for a possible IPO [15].

02 · Lesson · why it matters

Progress arrives as a demolition

The new thing creates value by destroying the old thing's - and the people it ruins are usually the ones who did everything right.

The week the frontier got cheap

Three of the largest AI companies cut prices this week, more or less at once. Anthropic shipped a model that does nearly what its best one does, for half the money. Microsoft published numbers claiming its own models undercut its partner OpenAI by up to 89%. Google put out a cheaper, faster model too.

This looks like generosity, or like ordinary competition. It is neither. It is a demolition, and it is worth learning to see it as one.

Value created is value destroyed somewhere else

There is a name for this, almost a hundred years old: creative destruction. The economist Joseph Schumpeter used it to describe how progress actually moves. Not by adding a new thing on top of the old ones. By replacing them - and wiping out what the old ones were worth.

The cheaper model does not sit politely beside the expensive one. It reaches over and takes the expensive one’s job. Every enterprise that switches to save money is money the frontier model no longer earns. The value the new model creates and the value it destroys are not two events. They are one event, seen from two chairs.

The loser did nothing wrong

Here is the part that stings, and the part worth carrying.

Last quarter’s top model cost a fortune to build. It works. It did not break, get hacked, or fall behind on quality. It is simply no longer the cheapest way to get good-enough work done - and that alone is enough to strand it. The company that built it did everything right. It aimed at the frontier, hit it, and shipped. Then a cheaper option appeared, and the frontier stopped being where the money was.

This is not failure. Failure is when you do the wrong thing. This is the harder case: you do the right thing, and the ground moves anyway. Most ruin in a fast-moving field is this kind. Not incompetence - obsolescence. The healthy incumbent, competent to the last, walking off a cliff it could not see.

It does not respect loyalty, or your own past work

Notice who is doing the demolishing. Microsoft has spent billions backing OpenAI. This week it published numbers showing its own homegrown models cost a fraction of OpenAI’s. The patron is quietly building the replacement for its own supplier.

And a lab’s newest model destroys the value of the one it shipped two months ago - its own, not a rival’s. The demolition does not care what you spent, who your partner is, or how recently you were ahead. Sunk cost is not a shield. The best position last quarter is just the tallest thing standing when the new wave comes.

You are standing in the demolition, on both sides

It is tempting to watch this from above, as a story about companies. It isn’t. You are inside it.

Your AI tools got cheaper this week. That is the good side of the same event that is stranding a vendor’s expensive model. The bill that fell on your desk is the margin that vanished from theirs. You benefit from a demolition. Someone paid for that.

And the mechanism does not stay in one industry. The skill a person spent ten years mastering is, to a machine, a task to be done cheaply. When the cheap-good-enough version of that skill arrives, the person did nothing wrong either. They got good at the right thing, and the ground moved. The same force that lowered your bill this week is the one that will, on some other week, come for a thing you built. Progress does not sort people into deserving and undeserving before it moves.

What looks permanent is usually a perch

For three years, the frontier lab’s lead looked like a fact about the world - a moat, a permanent advantage, the natural order of the market. It was never that. It was a temporary arrangement that happened to hold for a while, and posed as more.

Whenever an advantage looks permanent - a dominant chip, a default everyone uses, a skill that has always paid - it is worth asking a quiet question. Is this the shape of the world, or is it a perch that has held long enough to feel like the ground? The moat was real. It was also always going to be crossed.

The whole, held loosely

Creative destruction is not a glitch in progress. It is what progress is. We do not usually climb by stacking new things on old ones. We climb by knocking the old ones down and building on the rubble - and calling the rubble, once, the state of the art.

The uncomfortable truth inside that is simple. No seat can see when its own ground is next. The winners cutting prices this week are the incumbents someone will circle next year. The advantage that feels like bedrock is dated; you just don’t know the date. Watching a healthy giant get quietly displaced is not a story about their mistake. It is a rehearsal for a position every one of us will one day be standing in. It is also a reason to hold whatever ground we hold a little more loosely.

03 · Lab · your turn

The Incumbent's Dilemma

Rehearse the creative-destruction choice: protect your old margins and lose the market, ship the cheaper thing yourself, or wait and lose both.

04 · Hope · carry this

The same force that strands a costly model is the one that just cut your bill in half - progress mostly shows up as things getting cheaper and reaching more hands, and this week it reached yours.