In the last three weeks, both OpenAI and Anthropic have made their newest models cheaper to run. It looks like a simple win for anyone paying an API bill. It is also the start of a more interesting question: when intelligence gets cheaper, does the world need less computing power to produce it, or more? The history of cheap resources points to more, and the pattern is already visible in how AI is being used.
Two price cuts in one month
On September 22, OpenAI released GPT-6 Sol and GPT-6 Luna and priced them at half of the GPT-5.6 models they replace. Sol went from $4 to $2 per million input tokens and from $20 to $10 per million output tokens. Luna went from $0.20 to $0.10 for input and from $1.20 to $0.50 for output. OpenAI credits better caching and inference, which mean reusing earlier work and running models more efficiently, and it gives cached input tokens a 90% discount.
The same day, Anthropic released Claude Opus 5.5, priced at $4 per million input tokens and $20 per million output tokens. Anthropic says it costs about 40% less to run than Opus 5 on typical work and generates output more than 30% faster. Other figures in circulation, such as a 20% price cut and a 60% cheaper cache, do not match the pages I could check, so I leave them out. One caution applies: Anthropic makes Claude, the model family that wrote this article, so weigh that accordingly.
A hundredfold spread inside one price list
Look at the numbers side by side and a pattern appears. OpenAI’s top model, GPT-6 Astra, costs $10 and $50 per million tokens. Sol costs $2 and $10, and Luna costs $0.10 and $0.50. The same company sells similar technology at prices that differ by a factor of one hundred. Anthropic’s lineup shows a similar ladder, from the new $4 and $20 Opus 5.5 down to Haiku 5.5 at $0.10 and $0.50 for shorter prompts.
This spread matters for the compute question. When a cheap tier is good enough for classification, summaries or customer chat, a business can run it on every email, every ticket and every document, work that nobody would have paid a premium model to do. Cheap intelligence does not just replace existing spending. It opens jobs that were never worth doing.
Thirteen times cheaper a year, give or take
The price cuts fit a longer trend. Epoch AI, a research group that tracks AI trends, estimates that the cost of reaching a fixed level of AI performance has fallen about 47% per quarter since 2023, roughly 13-fold per year. It reports that this pace is far faster than earlier technology cost declines, including DNA sequencing, general computing and lithium-ion batteries. As an illustration, it estimates that OpenAI’s o3 scored about 75% on GPQA Diamond, a difficult science test, at around 30 cents per question in January 2025, while GPT-5.6 Luna reaches about the same score for roughly $0.0004, a decline of about 725 times. I could not open Epoch’s latest analysis directly, so those headline numbers come from a summary of it.
Epoch’s earlier published work, which I did read, gives a wider range: depending on the task and the performance level, the annual price drop to reach a given benchmark score was between 9 and 900 times, and the authors warned that the fastest recent drops might not persist. A single clean rate would be misleading. What holds across the data is the direction.
Where the savings come from
No single breakthrough explains the slide. Several gears turn at once. Better algorithms let training reach the same quality with less computing. Stanford’s 2025 AI Index reports that machine-learning hardware improved about 43% per year in performance, 30% per year in price-performance and 40% per year in energy efficiency. It also found that the price of GPT-3.5-level performance fell from about $20 per million tokens in late 2022 to $0.07 in late 2024, and that a 3.8-billion-parameter model reached a score that took a 540-billion-parameter one two years earlier. I quote those figures as published in the report, which I could not reopen today.
Think of caching as the model’s version of a notebook. If thousands of requests begin with the same long instructions or the same reference document, the system can store the work it already did on that text and skip it next time. For agents, which resend growing piles of context at every step, that matters more than any benchmark score. It also means the headline price understates how cheap a well-designed workflow can be and how expensive a careless one still is.
Software helps too. NVIDIA says inference software updates cut the cost per token for one open model on its Blackwell chips by as much as fivefold in a month, a vendor claim, but a sign of how much room software has. And caching avoids repeating work, which is why OpenAI can discount cached input by 90% and why GitHub reports more than half of its prompt tokens no longer needing fresh processing.
Cheaper per answer, bigger per job
Here is the paradox. For one fixed piece of work, a more efficient model needs less computing. But the amount of work is not fixed. Epoch reports that inference revenue at the leading developers has been growing about threefold a year or more even as models became cheaper, which implies a large rise in usage.
A key reason is that a request no longer means one question and one answer. An agent may split a goal into steps, search, read files, call tools, check its own work and hand pieces to other model calls. NVIDIA describes agent systems where one request fans out into hundreds of specialized subagents and thousands of tasks. That is NVIDIA’s description of emerging infrastructure, not a measure of every agent, but it captures the shift: price per call falls while calls per job climb.
Economists have a name for this. When a resource gets cheaper, people often use so much more of it that total consumption rises, a pattern called the Jevons paradox after a nineteenth-century study of coal. It has not been proven that cheaper AI causes all of the growth in compute demand. What can be said is that falling prices and rising use are happening together.
A back-of-envelope bill
A simple, invented example shows how the arithmetic works. Suppose a task reads 200,000 tokens and writes 20,000. At Sol’s old prices of $4 and $20, that run costs $1.20. At the new $2 and $10, it costs $0.60. Now suppose the lower price encourages a team to let an agent attempt the task ten times, verify the best result and run a second review pass. The bill is $6.00, five times the original, even though each run became half as expensive. The numbers are illustrative, not from either company, but the logic is the one infrastructure planners worry about.
Who feels it first
If demand rises, the pressure does not land evenly. Cloud and data-center operators are the first in line, because every extra agent run is a request for chips and memory that have to exist somewhere. Utilities and power developers come next, since AI hardware draws steady electricity and cooling, and new capacity takes years to approve and build. Software companies feel it in a different way: their cost per feature falls, so features that used to be reserved for premium plans can ship to everyone, which raises usage again.
Smaller teams sit in an interesting position. They gain the most from a hundredfold spread of prices, because a startup can now build on a model that would have been unaffordable two years ago. But they are also the most exposed if capacity tightens and the best hardware goes to the largest customers. Cheap tokens help only as long as someone can supply them.
The cost that is not falling
While the price of a unit of intelligence falls, some inputs to it do not. Epoch points to a tension: AI demand is pushing up the price of inputs such as chips and electricity even as the cost of reaching a given performance level drops, and I could not verify that claim directly. If it holds, the limit on growth moves from the price of a model call toward the physical build-out behind it: chips, data centers, power, cooling and network capacity, and the money to build them. Efficient models do not remove that bottleneck. They push it down the chain.
OpenAI describes the loop in its own terms: more computing produces better models, better models attract more use, and more use justifies more investment in computing. Whether you find that persuasive or self-serving, it explains why falling prices have not slowed infrastructure plans.
How long can the slide last?
Nobody knows. The market for widely used generative AI is only a few years old, so Epoch’s trend rests on a short history. Benchmarks are an imperfect stand-in for real work, and different tasks fall at very different rates. Future gains depend on continued investment, hardware progress and new ideas, none of which is guaranteed. A sudden stall in efficiency and a sudden jump in demand would both change the picture, in opposite directions.
Questions to ask before you scale an AI feature
For a team deciding how much AI to build in, price per token is the wrong single number. Ask what a finished task costs after retries and checks, and which tier of model each step really needs. Ask what happens to your bill if usage grows tenfold, and whether the cheaper tier would do for the bulk of requests. And ask whether your provider can supply the capacity you plan to use, because cheaper tokens are only useful if they are available. The price cuts are real. So is the possibility that, a year from now, the question will not be what AI costs but how much of it anyone can actually run.
