Pricing on a Falling Curve: AI Inference Costs, a16z’s 10x and Meta’s Muse Bet

a16z says the cost of equivalent LLM performance falls 10x a year, Epoch AI measured 9x to 900x, and Meta is counting on a more than 10x efficiency gain for Muse. Here is how founders can price against falling inference costs without betting the company on them.

On Oct. 21, 2026, Mark Zuckerberg told Meta’s investors that Muse “is very compute intensive to run” but that Meta expects “to improve its cost efficiency by more than 10x in the next year.” It is a high-profile bet on AI inference costs falling fast enough to turn an expensive free product into a profitable one. Founders face the same question at smaller scale: how far ahead of today’s costs can you set your prices?

Where it stands: The 10x figure is a Meta goal, not a result. CFO Susan Li said in the same call that “we expect total serving costs to rise” as Muse adds users, and Meta has not disclosed what Muse currently costs to run.

Why are AI inference costs falling?

Inference is the work a trained model does each time it answers a request. Outside researchers have tracked how quickly the market price of that work has dropped, and the numbers are large.

Andreessen Horowitz partner Guido Appenzeller put a number on it in November 2024: “For an LLM of equivalent performance, the cost is decreasing by 10x every year.” He illustrated it with history, noting that when GPT-3 became public in late 2021 it reached its benchmark score “at a cost of $60 per million tokens.” Tokens are the word-fragment units a model processes, and providers quote prices per million of them.

Epoch AI, a research group that tracks AI trends, found a wider spread. Prices fell “between 9x per year and 900x per year, with a median of 50x per year,” depending on the task and the performance level measured.

Source Measured Rate cited
a16z (Nov. 2024) Cost of equal performance 10x per year
Epoch AI (2025) Prices by task and threshold 9x to 900x per year
Meta (Oct. 2026) Muse compute cost efficiency More than 10x target

An important caveat: both outside studies measure market prices for a fixed level of capability. They do not measure what it costs Meta, or any single company, to serve its own product.

How does Meta plan to cut Muse’s costs by 10x?

Li named three levers on the call: “we expect to improve compute cost efficiency by more than 10x over the next year driven by better utilization of off peak capacity, more efficient models, and inference optimization.”

Each lever has a startup equivalent. Off-peak capacity means running work that can wait, such as background agent tasks, when servers are idle. More efficient models means routing simple jobs to cheaper models. Inference optimization covers techniques like caching and batching that squeeze more answers out of the same hardware.

Meta has a track record to point to on a different workload. Li said Meta’s own models were “reducing per unit content processing costs by more than 80% across Instagram threads, Facebook, and ads.” That is internal content processing, not Muse serving, but it shows the company has cut unit costs before.

The product shape makes the goal matter. Meta’s launch post says Muse “runs on its own dedicated computer in the cloud,” which is a heavier load per user than a chatbot.

What happens if the cost curve does not cooperate?

Zuckerberg’s profit case is conditional. In his words: “If compute efficiency continues improving quickly for another few years, which we think is highly likely, then Muse will be highly profitable over the years.” The word “if” carries the plan.

Wall Street noticed. Before the results, Bank of America analysts wrote, as quoted by CNBC, that Muse “will consume capacity w/o near-term revs.” And Li added that “third party AI token costs” were among the drivers of Meta’s expense growth, alongside legal and infrastructure costs.

The a16z post carries its own warning: “Given the early stage of the industry, the time scale may still change.” A trend that held for a few years is not a contract.

For a startup, the risk is sharper because there is no ad business to absorb a miss. OpenAI’s experience is the cautionary tale. In January 2025, CEO Sam Altman said the company was “losing money on its $200-per-month ChatGPT Pro plan because people are using it more than the company expected.” “I personally chose the price,” he wrote, “and thought we would make some money.” Customer usage outran the price he had set.

How should founders price when AI inference costs are falling?

Treat falling costs as upside, not as the base case. A few practical rules follow from the evidence above.

Price for today’s cost, plan for tomorrow’s. Set paid plans so they cover current serving costs at realistic heavy usage. When costs fall, raise allowances or cut prices. Customers welcome more for the same price; they resent less.

Track your own cost per task, not the market average. The a16z and Epoch figures describe list prices across the industry. Your product may use longer contexts, agents that run in loops, or premium models that fall in price more slowly. Our guide to calculating margin after inference costs shows how to measure it.

Build switching into the product. Meta’s second lever, more efficient models, only works if your system can route work between models. Design for that early.

Watch price signals from model vendors. Meta’s own API lists Muse Spark at $1.25 per million input tokens and $4.25 per million output tokens in Standard mode, and offers a much cheaper Contributor mode “in exchange for permission to use your content to train future Meta models.” Vendors trade price for data or commitments; read the terms before building on the lowest rate.

Where does falling cost fit in a pricing strategy?

It decides how generous you can be over time, not where you start. Meta can launch with a large free tier because, in Zuckerberg’s framing, a 10x efficiency gain is expected to “enable it to be profitable once the revenue lines are at scale.” A startup should reverse the order: earn margin first, then share the savings as costs come down.

If you do lean on falling costs, put a date on the bet. Write down the cost per task you need, the date you need it by, and what you will change in pricing if it does not arrive. That turns a hope into a plan you can check.

For related reading, our sibling articles cover usage-based pricing lessons and freemium when compute is expensive.

Epoch AI’s own range is the reminder that averages hide a lot: depending on the task, prices fell anywhere “from 9x to 900x per year.”

Frequently Asked Questions

Are AI inference costs falling?

Market prices have been. Andreessen Horowitz estimated in November 2024 that the cost of an LLM of equivalent performance was falling 10x every year, and Epoch AI found price declines of 9x to 900x per year depending on the task. Those are prices for a fixed level of capability, not any one company’s internal costs.

How much does Meta expect to cut Muse costs?

Meta said on its Oct. 21, 2026 earnings call that it aims for a more than 10x gain in Muse’s compute cost efficiency within a year, citing idle off-peak servers, leaner models and optimized inference. Meta presented it as an expectation; no result has been reported yet.

Should a startup price below cost and wait for inference to get cheaper?

Only with a plan for the case where savings arrive late. OpenAI’s CEO said in January 2025 that the company was losing money on its $200 ChatGPT Pro plan because people used it more than expected. Usage can grow faster than unit costs fall.

What is inference in AI?

Inference is the computation a trained model performs each time it answers a request. Training builds the model once; inference costs recur with every use, which makes it an ongoing cost for any AI product rather than a one-time expense.

Sources

This article is general information for founders and is not accounting, legal or investment advice. Figures are as reported by the sources cited on the dates shown; analyst and third-party estimates are not company disclosures. Found Founders is not affiliated with Meta.