The Short Answer
A million tokens — one "barrel of intelligence" — costs roughly $26 from one frontier lab, about $56 from another's newest model, and about $0.50 from Chinese providers, a spread of more than 50x for capability that now sits only 3–6 months apart. Spreads that wide do not persist, and no buyer can know whether they close by frontier prices falling or by subsidized inference being repriced upward.
The defensible position is therefore not a bet on which model wins. It is an architecture where the model is a swappable component and the layer above it — your agents, data, and workflows — is owned outright: you own all the code and the data.
That way, prices falling is a windfall you can capture in a configuration change, and prices rising is an inconvenience rather than a budget crisis.
What is the price spread in AI models right now?
Investor Chamath Palihapitiya frames it with a unit worth borrowing: a "barrel of intelligence," meaning one million tokens. His price comparison puts a barrel at about $26 from OpenAI, about $56 from Anthropic's most capable model, roughly $1 from xAI, about $1.50 from Meta, and about $0.50 from Chinese providers.
His summary of the arithmetic: you could buy 52 barrels of Chinese intelligence for the cost of one barrel from the premium tier.
Published rate cards show the same shape from a different angle. GPT-5.5 Pro lists at $30 per million input tokens and $180 output, while open-weight DeepSeek V4 Flash runs $0.054 input and $0.242 output — roughly 150x cheaper on output.
| Model | $/M input | $/M output | Capability index |
|---|---|---|---|
| GPT-5.5 Pro (closed) | $30.00 | $180.00 | Frontier |
| GLM 5.2 (open weight) | $0.447 | $3.31 | 51 |
| Nemotron 3 Ultra (open weight) | $0.423 | $2.61 | 48 |
| DeepSeek V4 Flash (open weight) | $0.054 | $0.242 | — |
Why hasn't the pricing gap closed with the capability gap?
Because the two gaps are driven by different things. Capability is a research race; price is a business model.
Palihapitiya's observation is that the capability gap narrowed much faster than the pricing gap — the capability difference is now modest while the price difference remains enormous.
The independent data supports the capability half. Open-weight models have held a consistent 3–6 month gap behind the frontier for over 18 months, with the frontier labs not visibly accelerating away.
Price, meanwhile, reflects what each provider needs to recover: enormous training runs and compute commitments in one case, strategic positioning or national industrial policy in another. Those motives can sustain a spread far longer than a pure efficiency market would.
They cannot sustain it forever, which is the part enterprises should plan around.
What happens to buyers when the spread rationalizes?
There are two directions, and the mistake is assuming only one is possible.
Closing downward. Competition and open-weight substitution drag frontier prices toward the floor. Good news for buyers — but only those positioned to actually move workloads and capture it. If your application is welded to one provider's API and behavior, you watch the price cut go to someone else.
Closing upward. Providers currently absorbing inference costs to buy market share stop doing so. Palihapitiya has argued the largest AI valuations could prove a "mathematical mistake", and repricing is the ordinary way that resolves. A buyer with no alternative absorbs it.
The asymmetry matters. In the downward case, lock-in costs you an opportunity. In the upward case, it costs you a budget you already committed to a board.
A three-year agreement priced against today's rate card is a bet that neither happens.
Does switching models actually work in practice?
It works when it was designed for, and it fails when it wasn't — which is why this is an architecture question rather than a procurement one.
What makes switching real: prompts and skills stored as portable assets rather than vendor-specific configuration; evaluations you can re-run against a candidate model to prove it clears your bar; routing that can send a workload to a different model without touching application code; and your data staying in your own systems throughout.
What makes switching theoretical: agent logic embedded in one vendor's orchestration product, no evaluation harness to prove a replacement is safe, and business data resident in the provider's environment.
The practical test is one question: how long would it take to move our highest-volume workflow to a different model, and how would we know it still worked? If the answer is a config change and a re-run of the evals, the spread is an opportunity. If it is a quarter of engineering, the spread is a liability already on your books.
What does a model-agnostic hedge look like?
It is the same posture regardless of which way prices move.
Own the layer above the model. Your agents, skills, ontology, and workflow logic are the accumulated asset. Keep them in a stack you control, so the model is rented and the intelligence you have built is owned — the argument we make in full in ownership versus rental economics.
Keep at least two viable models qualified at all times. Not aspirationally — actually evaluated against your workloads, with results recorded, so a switch is a decision rather than a project.
Route by task, not by default. Reserve the expensive tier for work that measurably needs it and let everything else run on the cheap tier.
Keep the option of running weights yourself. For regulated and sovereign workloads this is a compliance requirement anyway, and it converts a vendor price change into a hardware decision you control. We worked through the operational reality of this in the open-weight tipping point.
For U.S. institutions weighing counterparty risk alongside price, it is worth noting ibl.ai is family-owned and operated from New York, NY — the platform layer is not itself a bet on any one lab's business model surviving.
What should enterprises do in the next quarter?
Three concrete moves, none of which require predicting the market:
Price your switching cost. Put an actual number on moving your top workflow to another model. That number is your true exposure to the spread, and most organizations have never calculated it.
Qualify a cheaper model on a real workload. Take one high-volume, low-risk workflow and run it against an open-weight model with your own evaluation set. You either bank a large cost reduction or you learn precisely where the gap still bites.
Rewrite the procurement question. Stop asking only "which model performs best" and start asking "what does it cost us to change our mind, and who bears that cost." The second question is the one that determines what happens to you when the spread closes.
The spread will rationalize. Nobody gets to know in advance whether that arrives as a discount or an invoice. The only position that is comfortable either way is the one where switching is cheap — and switching is only cheap if you own everything above the model.