This week gave us two numbers that sit oddly next to each other. Anthropic and OpenAI both cut the price of frontier intelligence by around half. And SoftBank borrowed more than eleven billion dollars, at yields you would expect from a distressed borrower, to keep funding the machine that produces it. The output is getting cheaper for buyers. The capital that builds it is getting more expensive to raise. Anyone trying to understand where AI economics are heading should hold both facts in mind at once, because the gap between them is where the risk lives.
The price war nobody at the top wanted
Anthropic released Claude Opus 5.5 at four dollars per million input tokens and twenty per million output, and said typical workloads run about 40% cheaper than Opus 5 because the model burns fewer tokens and leans on cheaper cache reads. Roughly ninety minutes later OpenAI shipped GPT-6 Sol and Luna, with Sol priced at half of Opus and Luna priced low enough to be almost a rounding error for high-volume jobs. The timing was not a coincidence. When two labs launch within the same window and lead with the bill rather than a benchmark, they are telling you the market has moved from bragging rights to unit economics.
For buyers this is straightforwardly good. The same class of work costs meaningfully less than it did a month ago, and the practical question shifts from which model is smartest to which model clears the job at the lowest cost per successful task. That last phrase matters more than the sticker price. A cheaper model that needs more retries or more human rescues is not actually cheaper. Higher reasoning can buy quality but it also eats time, and time is a cost the pricing charts conveniently ignore.
There is a subtler point for the labs themselves. Cutting prices while your own cost of capital is climbing is a strange place to be. It works only if the buildout keeps driving the cost of serving each token down faster than competition drives the price down. That is a bet on engineering and scale, and it is not guaranteed to pay.
The plumbing, not the model, is where margins hide
The more useful cost story this week came from NVIDIA, which pointed a research AI at the harness around the model, the machinery that manages tools, context and files, and let it hunt for wasted tokens. Out of 152 candidate ideas, four survived. The result was roughly half the API cost of the native Codex and Claude Code setups at comparable quality, an estimated saving of nine to thirteen dollars per hour of agent work. The code is open and opt-in, and the patterns are worth stealing even if you never install it.
What the four fixes have in common is that none of them makes the model smarter. They stop it rereading a four-thousand-line log to use six lines, stop billing completed subtasks on every turn, stop replaying the same giant tool output. The waste was an engineering choice all along. For any company whose inference line grows faster than revenue, that reframes the problem. The cheapest model switch may not be a model switch at all. It may be cleaning up your own harness, and it is telling that a chip company chose to publish the playbook rather than sell you more chips.
The financing is where the strain shows
All of this cheaper output still rests on an expensive foundation. SoftBank raised $11.1 billion in what several outlets called the largest high-yield corporate bond sale on record, to fund the final $10 billion tranche of its $30 billion follow-on into OpenAI. The dollar tranches priced between roughly 8.6% and 9.75%, and the bonds carry BB+ ratings, which is to say speculative grade. That is nearly double what SoftBank paid on comparable debt a few years ago, and its credit default swap spreads have widened as investors ask harder questions about leverage.
The equity side looks healthier but no less frantic. Crusoe raised $3.9 billion at a $30.9 billion valuation, more than triple its worth ten months earlier, backed by Nvidia, sovereign wealth funds and a long list of blue-chip investors. The neocloud operators with the most contracted revenue can borrow against it cheaply through project finance, which is why the best-capitalised ones are pulling away. SemiAnalysis, which rates these providers, noted that only nineteen neoclouds worldwide now clear its quality bar, and it added a tier that politely describes the ones doing the bare minimum to get by.
So what
Cheaper tokens and pricier capital are two sides of the same buildout. As long as the scale race keeps pushing serving costs down faster than prices and financing costs rise, the model holds. The thing to watch is not the next benchmark or the next price cut. It is whether sentiment toward OpenAI and its peers stays firm, because a highly leveraged financing chain built on speculative-grade debt is only comfortable while the demand story holds. If that wobbles, the falling prices buyers are enjoying today will look a lot more fragile than they do right now.