LLMs Got Up to 10x Cheaper in a Year. Your AI Agents Did Not.

The price of intelligence is falling faster than almost anything else in software. For a fixed level of capability, the cost of running a model drops by roughly 10x a year, and by some measures far more. That is real, it is measured, and it is one of the best things to happen to anyone building on top of these models.

It is also where a lot of teams are about to make an expensive mistake with their AI agents. "The model got cheaper" and "my agent got cheaper" are not the same sentence, and the gap between them is where the bill quietly comes back.

The price collapse is real, and worth being clear about

Start with the good news, because it is genuinely good. a16z's analysis of inference costs found that for a model of equivalent performance, the price falls about 10x every year. The capability that cost 60 dollars per million tokens from GPT-3 in late 2021 cost about 6 cents from a small Llama model three years later, a thousandfold drop.

Independent measurement from Epoch AI backs the direction and often exceeds it. For a fixed benchmark score, they find prices falling anywhere from 9x to hundreds of times per year, with the steepest drops in the most recent year. If your job is to run last year's capability, it keeps getting cheaper on a schedule you can almost plan around.

That is a true Pareto improvement, and it deserves the enthusiasm before we complicate it.

Why "cheaper for the same quality" is not "cheap equals frontier"

Here is the distinction the enthusiasm usually skips. Those curves measure a fixed level of capability getting cheaper over time. They do not say that today's cheap model is as good as today's frontier model. Those are different claims, and treating them as one is the trap.

A 2025 study of what actually drives the declines put the frontier number more conservatively, around 5x to 10x a year once you separate it from hardware and competition effects, and found that the cost of running the actual frontier models is rising, not falling, as they grow larger and reason for longer. Its warning, in spirit: benchmarks can paint a warped picture of practical capability per dollar.

So a cheap model is cheap for a reason. It is usually a smaller, older, or more compressed model sitting behind the frontier, not the frontier at a discount. For plenty of work that is completely fine. For an agent, "plenty of work" is carrying a lot of weight in that sentence.

Where the cost goes when you buy the cheap model

Cheapness is manufactured, and each method of making a model cheaper trades away something specific. For a single chat reply you rarely notice. For an agent running thousands of steps, the thing that was traded away compounds.

None of these are hypothetical. Developers running Meta's cheap Muse Spark model this month reported it getting stuck re-reading and rewriting files it had just written, and calling tasks finished when they were not. The long-context research is blunter still: models do not use their windows uniformly, accuracy drops for information in the middle, and effective context can fall far short of the advertised number. A cheap "long-context" tier may not actually hold a long agent trajectory.

The only cost number that tells the truth for agents

Per-token price is the wrong scoreboard for an agent, and it is the one everyone watches. The honest metric is cost per successfully completed task.

The difference is retries. When a cheap model loops, misfires a tool call, or produces an output that a check rejects, the agent runs the step again, often re-sending the whole context, and the cost of that one task multiplies. Two retries can triple the cost of a session. A model that is cheaper per token but needs four attempts is not cheaper, it is four times the work for the same result.

This is why a more capable, more expensive model is often the cheaper one end to end. The research is careful here rather than triumphant: stronger models mainly reduce iteration overhead, finishing in fewer loops, which offsets their higher per-token price rather than beating it outright. The point is not that premium always wins. It is that per-token price alone cannot tell you which model is cheaper, and for agents it usually misleads.

The discount that is really a data trade

One more cost hides in plain sight, and it is not paid in tokens. The cheapest tier of a model is sometimes cheap because you are paying with your data.

Meta's Muse Spark is the clearest current example. Its Contributor tier is about twelve times cheaper on input than the standard tier, and the reason is stated on Meta's own pricing page: Contributor inputs and outputs are used to improve Meta's products, standard-tier data is not. That is a fair option for public or synthetic work. It is the wrong default for anything regulated or confidential, and the low price is exactly what makes it tempting to reach for without reading the terms.

How to capture the savings without buying the downgrade

The falling price is real money and you should take it. The way to take it safely is to stop buying one model for the whole agent and route by the job instead.

The savings are large when you do. Cost-optimization teams report cutting agent costs by 40 to 70 percent by moving the majority of requests to cheaper models with no perceptible quality loss, precisely because they keep the cheap model on the high-volume, low-stakes, short-horizon work and reserve a frontier model for the steps that carry consequences or need real reasoning. The savings come from routing, not from blanket cheap-model selection.

That is how we think about it at Beam. The model is set per step, cheap and fast where the work is simple and repetitive, stronger where a mistake is expensive, with fallback and automatic re-runs catching the failures that come with running many models. The cheaper models get, the more of an agent they can safely cover, as long as the layer around them measures cost per completed task rather than cost per token. If you are deciding which model belongs on which step, we wrote a separate piece on assigning models per step that goes deeper.

The price of intelligence will keep falling. The teams that come out ahead are the ones who spend the savings on reliability instead of assuming the discount arrived for free.

Common questions about cheaper LLMs and AI agent costs

Are cheaper LLMs worth it for AI agents?

Often yes, for the right steps. Cheaper models are a strong fit for high-volume, low-stakes, short-horizon work, and routing that work to them can cut agent costs by 40 to 70 percent. They are the wrong choice for long-horizon reasoning and high-stakes steps, where the retries they trigger cost more than the frontier model would have.

Does a lower price per token mean a lower cost per task?

No. If a cheap model needs retries, loops, or human rework, the cost per successfully completed task can be higher than a more capable model that finishes on the first attempt. For agents, measure cost per completed task, not per token.

Is model quality dropping as prices fall?

Two different things are happening. A fixed level of capability gets cheaper every year, which is a real gain. But choosing a cheaper model today usually means a genuine step down from today's frontier, because that model is smaller, older, or more compressed. The price curve is not evidence that the cheap model is as good.

Why is one model cheaper than another?

Sometimes through real efficiency, such as better architectures, near-lossless FP8 quantization, or caching, and sometimes through trade-offs that hurt agents, such as aggressive INT4 quantization, distillation, or fewer reasoning tokens. And occasionally the discount is a data trade, where a low-cost tier trains on your inputs and outputs.