GPT-4o mini costs $0.15 per million input tokens. GPT-3.5-Turbo cost $0.50. Same company, roughly eighteen months apart, and the floor moved about 70% — while the hardware underneath it moved maybe 20%.
That gap is the story. Not the press release. Not the "aggressive push against open source." The distance between the price decline and the hardware cost decline is where every lazy narrative about this event gets built.
I pulled the pricing pages again last week — OpenAI's, Anthropic's, Google's — and set them beside the amortization spreadsheets I built in 2020 for the Uniswap V2 pool I was running. That pool taught me more about marginal cost than any finance course I ever sat through. The number that matters here is not $0.15. The number that matters is the derivative. The slope of the cost curve is steeper than the slope of the hardware curve, and that spread is being paid for by somebody's balance sheet.
Consensus is broken. The market read these cuts as a product war. They are a capital event wearing a product war's clothing.
Let me set the bones.
Sometime in the late summer of 2024 — the window is August into September — OpenAI repriced GPT-4o mini and cut the GPT-4o rate. Anthropic had already parked Claude 3 Haiku at $0.25 input and $1.25 output per million tokens, and spent the quarter optimizing the middle of its stack. Meta, back in July, pushed Llama 3.1 405B into the field with the usual "open-source frontier" framing, a marketing posture more than an economic one. Google's Gemini 1.5 Flash was already sitting below the ten-cent input threshold, subsidized by TPU economics that neither OpenAI nor Anthropic can replicate at the silicon level.
The popular frame: two closed-source labs are cutting prices to strangle open-source models. Llama, Mistral, the whole menagerie.
That frame is lazy, and it is wrong in a way that matters.
Here is what the API actually is. It is a metered pipe. Every token in and out carries a real marginal cost, dominated by three things: GPU time, memory bandwidth, and the electricity that cools the rack. When I modeled a CBDC settlement pilot's inference load in 2024, the honest per-million-token cost of a mid-tier model on rented H100 capacity ran somewhere between the published API price and twice it, depending on utilization. That utilization number is the whole game. Idle GPUs are the most expensive object in this industry. The price cuts are not generosity. They are a utilization strategy dressed as competition.
Now the capital structure. Anthropic has raised north of ten billion dollars, including roughly four billion from Amazon and commitments from Google in the two-billion range. OpenAI sits above fourteen billion, with a thirteen-billion-dollar compute credit line from Microsoft. These are not companies optimizing for margin. They are companies optimizing for installed base, and they are doing it with other people's money.
Read the price cuts again. They are not a discount. They are a liquidity subsidy, drawn against the next round.
This is where the mechanics get interesting, so let me slow down.
The inference cost collapse has three engines, and only one of them is hardware.
The architectural engine is the one I find most elegant. The shift from dense transformer inference to sparsely activated mixture-of-experts changed the ratio between parameters stored and parameters computed. A model can carry trillions of parameters while touching a small fraction per token. That decouples capability from per-token compute in a way the 2017-era Ethereum gas debates I was living inside could not have anticipated. Back then I wrote a fifteen-page internal memo to my firm arguing that the block gas limit controversy was being misframed — the bottleneck was never block size, it was computational complexity. I was ignored. Same lesson here, eight years later. The bottleneck is never the raw number. It is the efficiency of the path.
The runtime engine is where the real money sits. Speculative decoding, prefix caching, KV-cache reuse across requests, dynamic batching with high fill rates. Each of these is worth somewhere between 1.5x and 3x throughput. Stack them and you get a compounding reduction in cost per token that has nothing to do with NVIDIA's price list. This is production infrastructure that matured inside eighteen months. When I ran my Curve Finance stability case study after the 2020 yield farming experiment, the thing that struck me was how much of the protocol's apparent safety came from mechanism design rather than collateral. Same pattern. The safety of an API margin comes from engineering, not from hardware subsidy.
The scale engine is the oldest law in commodity markets. When daily inference volume crosses into the trillions of tokens, cross-request cache reuse becomes statistically significant. The larger the batch, the better the fill, the lower the marginal cost. This is why the price cut is defensible at all. It is not charity. It is the flywheel spinning.
Now the macro layer, because this is where crypto people should be paying attention and mostly aren't.
The API price collapse is downstream of the dollar liquidity cycle, and that cycle is turning.
I have done this work before. In 2022 I reverse-engineered the Terra death spiral and mapped it against global M2 expansion. The finding then was that LUNA was not a crypto-native failure. It was a leveraged bet on the continuation of a dollar liquidity flood, and it died when the flood receded. The mechanism was an old one: liabilities marked to a floating rate, assets marked to a narrative. Terra was a duration mismatch wearing a stablecoin costume.
The AI capex boom has the same skeleton. Every training run, every inference cluster, every subsidized token is financed by cheap capital expecting a future payoff. Anthropic burns against AWS credits. OpenAI burns against Microsoft's balance sheet. The price cuts accelerate the burn in exchange for market share that is supposed to justify the next raise. This works beautifully while capital is cheap and equity markets keep repricing AI upward. When that changes — and I am not predicting when, I am describing the sensitivity — the subsidy model cracks. The first signal will not be in the API price. It will be in the financing terms. Watch the debt tranches, not the pricing page.
I made this argument in a 2024 report on what I called liquidity migration patterns, comparing the ten-billion-dollar institutional inflows into spot Bitcoin ETFs against the 2017 ICO era. The finding that stuck with me: the underlying protocol never changed, only the settlement layer's accessibility. Same lens applies here. The AI price war is not changing the nature of inference. It is changing who can afford the pipe. And the pipe always gets paid for by the same party — the one holding the dollar float.
Here is the crypto crossover, and it is brutal.
Render, Akash, io.net, and the rest of the decentralized-compute cohort built their entire valuation on a single claim: that distributed, permissionless GPU supply can undercut the hyperscaler. For years that claim held, because centralized inference was expensive and the networks could price twenty or thirty percent below AWS on spot-equivalent terms.
That gap just closed. Or worse than closed — it inverted.
A startup choosing between running Llama 3.1 70B on an Akash lease and calling GPT-4o mini at fifteen cents per million input tokens faces a decision that is no longer close. The self-hosted path carries GPU amortization, orchestration overhead, ops labor, model refresh, and the operational risk of a cluster going dark at 3 a.m. The API path carries a metered bill and nothing else. At low and moderate volumes, the API wins on total cost of ownership — and it wins on capability too, because GPT-4o mini outperforms a 70B Llama on most benchmarks that matter.
Yields are traps. The staking yields on AI DePIN tokens are not compensation for useful work. They are emissions funding a narrative that the underlying compute is economically necessary. When the centralized price floor drops below the decentralized marginal cost, the emissions have nothing to attach to. The yield keeps paying. The work doesn't arrive.
I have watched this pattern before. In 2021 I ran an audit with three junior analysts across fifty NFT collections, stress-testing interoperability claims. Four percent held up. The rest were ownership theater — claims of digital scarcity built on data layers that didn't connect to anything. The report was dismissed as bearish noise. It aged fine. The same forensic lens applies here. Pull the cost-per-token disclosure from any decentralized compute network and compare it to GPT-4o mini. Not the marketing number. The audited, load-adjusted number. The gap is wider than the networks admit, and it is widening.
Scale kills decentralization. This is not ideology. It is arithmetic. Inference has a utilization curve, and decentralized networks structurally cannot fill it. A permissionless cluster cannot guarantee the 90%-plus GPU fill rates that let a centralized provider offer fifteen cents. The physical constraints — bandwidth between nodes, coordination latency, the cost of redundant compute — are not governance problems you can vote away. They are physics.
So what is happening underneath the price war is a layer collapse. The centralized labs and the hyperscalers are merging into a single pricing bloc. Microsoft Azure plus OpenAI. AWS plus Anthropic. Google Cloud plus Gemini. Three vertical stacks, each with its own silicon, its own model, its own distribution. Open source is the fourth bloc, and its entry ticket is distribution, not economics. Meta gives Llama away because Meta monetizes attention, not tokens. Llama is a strategic tax on the closed-source labs, not a business.
There is another casualty nobody is naming. The white-label API resellers — the middle layer that bought tokens in bulk and resold them into vertical markets — just got washed out. Their entire margin was the spread between wholesale and retail API pricing. When the wholesale price falls 70%, the retail spread compresses toward zero. That layer dies quietly, and most of its operators are small enough that no one writes a post-mortem.
And then the model-layer startups. The Coheres, the Mistrals, the Together AIs. Their equity narrative was "efficient inference on open weights." When a frontier lab offers better capability at the same price point, the differentiator evaporates. The valuation reset has not fully happened. It will.
Here is where I break from the crowd, including most of the people who agree with the paragraph above.
The consensus is that this is closed source versus open source. Tidy. And wrong.
The real war is OpenAI against Anthropic against Google, with each one trying to become the default substrate of the AI economy. Open source is the narrative wallpaper. Look at actual behavior. Every price cut lands in the small and mid tiers — the mini models, the Haiku models — not the flagships. That is deliberate stratification. The labs want high-volume, low-complexity, high-frequency tasks to land on cheap models where they can kill competition on unit economics, while keeping the flag models at premium prices where margin lives. It is a pyramid, and they are defending the apex.
If the target were really open source, you would see them attack the fine-tuning market aggressively, or open the flagship tiers. They haven't. They defended the flagship price band and carpet-bombed the floor.
The second blind spot is Google. Gemini 1.5 Flash is already below the ten-cent threshold, and Google owns the silicon — TPU capacity that does not transit NVIDIA's pricing. Neither OpenAI nor Anthropic enjoys that. Microsoft and Amazon are landlords who can raise the rent. Google is a landlord to itself. Every dollar of API margin Google gives up, it partly recovers on TPU utilization and Workspace and Android distribution. That is a structurally cheaper position than either of the two labs that just cut prices, and almost nobody is modeling it as the actual long-term threat.
The third blind spot is the one that should unsettle the crypto reader. The squeeze on open source is being read by the crypto market as bullish for decentralized AI. It is the opposite. The cheaper centralized inference becomes, the harder the "decentralize the compute layer" thesis is to fund. The tokens that rallied on it were pricing a world where centralized inference stayed expensive. That world is closing.
So here is what I am actually tracking, and none of it is in the press releases.
The elasticity number. If the price cut is 70% and API volume growth is less than 300% on a comparable time base, the subsidy is not paying for itself and the next cut will be defensive rather than offensive. If volume growth clears 300%, the labs have found their flywheel and the squeeze continues. Watch quarterly disclosures, not announcements.
The crossover point on audited decentralized compute. The first network to publish a load-adjusted, third-party-verified cost-per-token below GPT-4o mini changes the conversation. If that number does not appear within eighteen months, the DePIN compute narrative has no floor underneath it, and the tokens should reprice.
The financing terms. Cheap dollar liquidity is the hidden input to every subsidized token. When capital markets stop funding the burn, the price war stops being a strategy and becomes a retreat.
The question worth sitting with is not who wins the price war. It is whether a market where the cost of cognition falls 70% every eighteen months can sustain more than three producers. Every commodity market in history has answered that question the same way. The only variable is the timeline.