To a large extent, it is a rapid decline in cost of the underlying commodity that drives technology cycles. Two prominent examples of this are Moore’s Law and Dennard scaling, which help to explain the PC revolution by describing how chips become more performant over time. A lesser known example is Edholm’s Law, which describes how network bandwidth increases — a key factor in the dotcom boom.
In analyzing historical price data since the public introduction of GPT-3, it appears that — at least so far — a similar law holds true for the cost of inference in large language models (LLMs). We’re calling this trend LLMflation, for the rapid increase in tokens you can obtain at a constant price.
In fact, the price decline in LLMs is even faster than that of compute cost during the PC revolution or bandwidth during the dotcom boom: For an LLM of equivalent performance, the cost is decreasing by 10x every year. Given the early stage of the industry, the time scale may still change. But the new use cases that open up from these lower price points indicate that the AI revolution will continue to yield major advances for quite a while.
The graph below shows the cost of the cheapest model for any month that would give us a minimum MMLU score of 42.
When GPT-3 became publicly accessible in November 2021, it was the only model that was able to achieve an MMLU of 42 — at a cost of $60 per million tokens. As of the time of writing, the cheapest model to achieve the same score was Llama 3.2 3B, from model-as-a-service provider Together.ai, at $0.06 per million tokens. The cost of LLM inference has dropped by a factor of 1,000 in 3 years.
If we pick a higher MMLU score of 83, we have less data because models of this quality level have only existed since GPT-4 came out in March of 2023. Since then, however, the price for models at this level have come down by about a factor of 62.
In the logarithmic plot below, we can see that the trend of a 10x decrease every year (the dashed line) is a fairly good approximation of the cost decline across both MMLU performance levels.
While we think the overall result is valid, the methodology is far from perfect. Models can be easily contaminated or intentionally trained on the MMLU benchmark. We also, in some cases, could only find multi-shot data for MMLU (although we are not including any chain-of-thought results in our data). And other models and fine-tunes may have been slightly more cost-effective at any given time. All that said, there is no question we are seeing an order of magnitude decline in cost every year.
Will LLM prices continue to decline at this rate?
This is very hard to predict. In the PC revolution, cost decreased to a large degree as a function of Moore’s Law and Dennard’s Law. It was easy to predict that as long as these laws held and transistor counts and frequencies increased, price drops would continue. In our case, however, the decrease in the cost of LLM inference is caused by a number of independent factors:
- Better cost/performance of the GPUs for the same operations. This is a result of Moore’s Law (i.e., the increasing number of transistors per chip), as well as structural improvements.
- Model quantization. Initially, inference was done at 16-bits, but for Blackwell GPUs we expect 4-bit to become common. That is a net increase of at least 4x in performance, but likely more as less data movement is required and arithmetic units are less complex.
- Software optimizations that reduce the amount of compute required and, equally importantly, reduce the required memory bandwidth. Memory bandwidth previously was a bottleneck.
- Smaller models. Today, we have a 1-billion-parameter model that exceeds the performance of a 175-billion-parameter model just 3 years ago. A major reason for this is training the models on a larger number of tokens, far beyond what was considered optimal based on Chinchilla scaling laws.
- Better instruction tuning. We have learned a lot about how to improve models after the pre-training phase, with techniques such as reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO).
- Open source. Meta, Mistral, and others have introduced open models that can be hosted by competing, low-cost model-as-a-service providers. This reduced profit margin across the value chain, which lowered prices.
There is no doubt we will see rapid advancements in some of the areas, but for others, like quantization, it is less clear. So while the cost of LLM inference will likely continue to decrease, its rate may slow down.
Another important question is whether this rapid decrease in cost is a problem for LLM providers. For now it seems like they are willing to concede the low end of the market and instead focus their efforts on the highest-quality tier. Interestingly enough, OpenAI’s leading model today, o1, has the same cost per output token as GPT-3 had at launch ($60 per million).
That said, the rapid decrease of LLM inference cost is still a massive boon for AI in general. Every time we decrease the cost of something by an order of magnitude, it opens up new use cases that previously were not commercially viable. For example, humans can speak around 10,000 words per hour. If someone were to speak for 10 hours a day, every day of the year, they could now use a GPT-3-class LLM to process all the words they said for about $2 per year. Processing the entire Linux kernel (about 40 million lines of code) would cost under $1.
Text-to-speech models are equally cheap, so building a simple voice assistant is now essentially free from an inference perspective.
The community will continue to build amazing applications around this technology, and we are super excited to partner with the founders who create the breakthrough companies that bring them to market. It is a great time to be an entrepreneur!
