When GPT-4 launched in March 2023, running it cost around $30 per million tokens. Today a frontier-class model costs roughly $2.50 for the same volume, and a perfectly capable budget model costs about $0.10 — three hundred times cheaper than GPT-4 was at launch, in a little over three years.

Sit with that number, because it’s the most important trend in the industry and the least discussed at dinner. The thing everyone is racing to build — raw machine intelligence, sold by the token — is collapsing in price faster than almost any technology in history. Across frontier models, average output prices have fallen roughly 94% since early 2023.

When the price of something falls like that while the quality keeps rising, you’re not looking at a product anymore. You’re looking at a commodity. And the strategic consequences of that are enormous, because in a commodity market the value doesn’t live in the commodity. It lives in what you build on top of it.

I see this from my own desk: I run Hermes, my personal agent, partly on open-weight models on a home server. Frontier-grade reasoning, no frontier-lab invoice. Three years ago that sentence would have been science fiction. That it’s now mundane is the story.

For the rest of us: what “commodity” means here

A commodity is a good that’s essentially the same no matter who produces it, so buyers choose on price and availability rather than brand. Electricity is the classic example. The electrons from one power station are identical to another’s; you don’t agonise over which utility’s electricity to run your fridge on. You care that it’s cheap, reliable, and there when you flip the switch.

The “token” is the unit AI models are sold by — roughly a word-piece of text in or out. So “$30 per million tokens” is just the price of a fixed chunk of machine thinking.

The claim of this piece is that machine thinking is on the same path as electricity. Not yet fully there, but heading there fast: the models from different labs are becoming close enough in quality, and cheap enough, that for most uses you’ll pick on price and convenience rather than logo. When that happens, owning the best model stops being a durable advantage — the same way no business today wins by owning a slightly better power socket.

The price collapse is the headline

The raw numbers are worth laying out, because the speed is genuinely hard to believe.

GPT-4 at launch (March 2023): about $30 per million tokens. By July 2024 — sixteen months later — GPT-4o mini delivered roughly GPT-4-class quality at $0.15 per million tokens. That’s a 200-fold drop in the price of a given level of intelligence, in not much more than a year. The cheapest capable models today sit around $0.10 per million. Even at the very top, where you’d expect prices to hold, frontier models have gotten about twelve times cheaper.

This isn’t a temporary price war that resolves into stable margins. It’s structural, driven by forces that all point the same way: GPU hardware gets several times more efficient each generation; inference software keeps finding optimisations; mixture-of-experts architectures activate only a fraction of a model’s parameters per token; open-weight competitors set a free-ish price floor; and the sheer scale of usage spreads fixed costs across billions of requests. Every one of those drivers is still accelerating.

Intelligence, in other words, is deflating. Plan as if a unit of it will cost a fraction next year of what it costs today, because it will.

The models are converging

Falling price would matter less if one lab stayed decisively ahead. It isn’t happening. The frontier is crowded and the pack is tight.

On most benchmarks, the leading models from OpenAI, Anthropic, Google, and a handful of others cluster within a narrow band — close enough that the “best” model changes month to month and rarely by a margin that decides a real product. More striking, open-weight models from DeepSeek, Meta, Alibaba and others now trail the closed frontier by months, not years. DeepSeek’s R1 in early 2025 made the point loudly: a follower can reach near-frontier capability fast, and give the weights away.

Convergence plus collapsing price is the exact recipe for commoditisation. When the products are nearly interchangeable and the price falls toward the cost of compute, there’s no premium left to defend. The model becomes a swappable input.

So where does the value go? Up the stack.

If the model is the commodity, the value migrates to the layers above it — the same way value in computing moved off the bare processor and into software, and value in electricity moved off the generator and into everything you plug in.

Above the model sits the application and, increasingly, the service: the company that uses cheap, abundant intelligence to actually do the work and sells the outcome. That’s the thesis I unpacked in The Next Trillion-Dollar Company Won’t Look Like Software — the value isn’t in selling access to a model, it’s in selling the finished result the model makes possible. Cheap models are the precondition for that whole shift. The economics of selling work at software margins only close when the underlying intelligence is nearly free.

The practical design pattern that follows: treat the model as interchangeable. The teams building well now put an abstraction layer between their product and any specific model, so they can swap providers as price and quality shift — exactly how you’d treat any commodity input. Your moat is your proprietary data, your workflow, your distribution, the trust you’ve earned with the buyer. It is emphatically not which model you called this quarter.

What this means

For anyone making AI decisions, commoditisation of the model layer changes the calculus in three concrete ways.

Don’t build your strategy on having the best model. You won’t, durably, and it won’t matter. The model you’re impressed by today will be cheap and matched within months. Build on the things that compound — data, workflow, distribution, domain trust.

Architect for swappability. Put a thin abstraction between your application and the model provider. Assume you’ll change providers more than once, chasing price and capability. Don’t hard-wire your business to one lab’s API any more than you’d wire your factory to one power company with no switch.

Budget for intelligence getting cheaper, not dearer. Use cases that look marginal at today’s prices may be obviously worth it next year. Plan capacity and roadmaps around a falling cost curve. The constraint is shifting from “can we afford the tokens” to “do we have the data and the workflow to put cheap intelligence to work.”

The labs building frontier models are doing something extraordinary, and they’ll remain enormous businesses. But the durable, defensible value of the AI era won’t accrue to whoever owns the best model this month. It’ll accrue to whoever builds the most valuable thing on top of a commodity everyone can buy. The interesting question is no longer “whose model is best.” It’s “what will you do now that genius is cheap.”

References

  • DeepLearning.AI, The Batch — “Falling LLM Token Prices and What They Mean for AI Companies.”
  • Public LLM-pricing histories (TokenCost, BenchLM, TokenMix) — GPT-4 ~$30/M (March 2023) → frontier ~$2.50/M and budget ~$0.10/M (2025–26); ~94% average frontier output-price decline since 2023; GPT-4o mini at $0.15/M (July 2024).
  • DeepSeek R1 technical report (January 2025) — open-weight near-frontier capability.
  • keller-ai — The Next Trillion-Dollar Company Won’t Look Like Software; Hermes, the Agent.