Half-price AI is the real frontier, not smarter models

76% lower token prices show AI competition has shifted from intelligence to cost.

AI is no longer being sold on the promise of being the smartest model in the room; it is being sold on being cheap enough to run all day inside agents.

Meta’s Muse Spark 1.1 and xAI’s Grok 4.5 are the clearest proof yet. Meta priced Muse Spark 1.1 at $1.25 per million input tokens and $4.25 per million output tokens, while xAI priced Grok 4.5 at $2 and $6. Those are not cosmetic discounts. They are strategic cuts aimed at the enterprise workloads that burn through tokens by the millions. When a model is invoked by software agents instead of humans, every cent matters, and the company with the lowest reliable price wins more deployments.

The market is already rewarding that logic. BigGo’s source notes that the average token price index fell 22% in a single month, from $2.10 per million tokens in May to $1.64 by July 7. That is the kind of move you see when buyers stop treating frontier AI as a prestige purchase and start treating it like cloud infrastructure. Once procurement teams compare models the way they compare storage or bandwidth, the vendor with the best benchmark score loses leverage to the vendor with the best unit economics.

First argument: agents turn token pricing into a board-level issue

AI agents change the economics of model choice. A chatbot answers once. An agent searches, drafts, retries, executes, summarizes, and loops until the task is done. Gartner’s forecast that enterprise software with AI agents will rise from under 5% last year to 40% by the end of this year means token consumption will explode. In that world, a model that is 60% or 76% cheaper is not a bargain. It is the difference between a feature that scales and one that gets turned off by finance.

This is why Meta’s language matters. Alexandr Wang called Muse Spark 1.1 a “workhorse” rather than a showpiece. That framing is the tell. The company is not trying to win a beauty contest on benchmark charts. It is trying to become the default engine for coding, task execution, and other long-running workflows where reliability and cost beat novelty. Meta knows that if it can own the cheap, competent layer, it can lock in developers before they ever feel the need to pay for premium intelligence.

Second argument: commoditization is forcing everyone down the stack

The pricing war is not happening in a vacuum. Chinese models such as DeepSeek, Zhipu AI, and MiniMax have already trained the market to expect 60% to 90% lower token prices than US frontier offerings. OpenRouter data showing Chinese models taking more than 30% of weekly token usage since February is the clearest signal that developers vote with their workloads, not with brand loyalty. Once low-cost models become acceptable for summarization, customer service, and data organization, the old premium tier starts to look inflated.

That pressure reaches even the most famous US labs. OpenAI’s move to split GPT-5.6 into three tiers, including a cheaper “Luna” option at $1 input and $6 output, reads as a defensive response. The pattern is unmistakable: the market is no longer asking which lab can produce the most impressive demo. It is asking which lab can deliver enough quality at a price that survives repeated use. Stanford’s AI Index noting that the performance gap between top models has narrowed to a razor-thin margin only sharpens that reality. When quality converges, price becomes the battlefield.

The counter-argument

The strongest objection is that cheap models are not enough for serious work. Enterprises still need reliability, safety, and deep reasoning, and the highest-end models remain the best choice for legal analysis, advanced coding, finance, and other high-stakes tasks. If a cheaper model produces errors that require human correction, the savings vanish. In that view, the race to the bottom risks encouraging vendors to underinvest in frontier progress and pack the market with fast, shallow systems.

That criticism is valid, but it does not overturn the argument. It proves the opposite: the market is splitting into tiers. The premium layer will survive for hard problems, while the volume layer will belong to cheap, competent models. That is exactly why pricing matters more now than raw intelligence. Most enterprise usage is not a heroic research problem. It is repetitive work at scale. For that layer, the best model is the one that is good enough and inexpensive enough to run continuously.

What to do with this

If you are an engineer, PM, or founder, stop evaluating AI as a single model purchase and start treating it as an operating-cost decision. Route simple tasks to low-cost models, reserve premium models for high-stakes reasoning, and measure total task cost rather than per-call novelty. The winners in this market will not be the teams that chase the smartest demo. They will be the teams that build systems where intelligence is allocated like compute, with discipline, not romance.