The cheapest model at GPT-4's level got 300x cheaper. The most capable model got 7.5x cheaper.
Understanding how AI model pricing has evolved over time can be instructive. I first explored this in November 2023. OpenAI had just announced new pricing for GPT-3.5 Turbo: $1 per million input tokens. For context, this was a 20x decrease compared to Davinci ($20) in March of that year.
If we revisit the data today (see chart below based on 192 models from 13 providers), the good news is that the price decrease trend still holds: GPT-4's capability cost $30.00 per million input tokens in March 2023 and $0.10 today, 300x less. But there is some clear separation.
There are now two categories of models - small but capable models, and large state of the art models. At the bottom are the mini and flash families. Of the 16 on sale with a score that was measured, 14 beat GPT-4, the best model of 2023, on the capability measures used here, at $0.10 to $1.50 per million input tokens. At the top are much bigger models trained on much more compute. The most capable one on sale costs $4.00 today, and averaged over each year, from GPT-4's launch in March 2023, it cost $26.18 in 2023, $14.14 in 2024, $7.02 in 2025 and $5.39 so far in 2026.
One interesting question we can answer from all of this is around if "AI is getting cheaper"?. The data suggests this is true but with some nuance. The small model regimes have gotten better and cheaper 300x and while state of the art or larger models have also gotten cheaper but by a much smaller factor 7.5x.**
The chart below draws two lines through the prices. The lower one follows the cheapest model on sale at GPT-4's level or better, where GPT-4's level is its score on the Artificial Analysis Intelligence Index, 6.7 against 57.6 for the most capable model today. The upper one follows the most capable model on sale.
Figure 1.
Each dot is one published price, and a short step joins a model's old price to its new one where it was repriced. The two lines are one line until March 2024, when Claude 3 Opus took the top score at $15.00, and they end 40x apart. The upper line dips wherever a cheaper model held the top score, and 18 of the 23 models that have held it have estimated scores. Try: pick a provider to see its models alone, or up to four to compare them. Switch the class to Small or Large to see each end of the market, and hover any point for its source.
Budget for the most capable model at today's price, and re-test cheaper models every few months.
In September 2025 the most capable model on sale was GPT-5 at $1.25. Today it is Claude Opus 5.5 at $4.00, and a model that scores at least as high as GPT-5 costs $0.10 (GPT-6 Luna).
The price of the most capable model on sale has changed 19 times since March 2023, each time because another model took the top score, and 11 of those changes were rises.
If your workload mixes easy and hard requests, send each request to the cheapest model that can do it. Microsoft Foundry's model router and Amazon Bedrock's intelligent prompt routing are managed versions of this.
This changes when a provider launches its next top model at a lower price. Claude Opus 4.5 launched at $5.00 in November 2025, against $15.00 for the Opus before it, and Claude Opus 5.5 took the top score in September 2026 at $4.00.
Limits: The figures in this section are prices per million tokens. One study measures the cost of running a whole benchmark, and finds that a fixed score got 5x to 10x cheaper per year while running the best model got 3x to 18x more expensive per year.
Models within one index point of each other can differ 33x in price
If we take every model on sale today with a measured score on the index, 42 pairs of them score within one index point of each other. In the median pair the more expensive model costs 2.5x the cheaper one. In the widest pair it costs 33x, with Claude Opus 4.8 at $5.00 and GLM-5.3-Flash at $0.15.
Within five index points of the most capable model, Claude Opus 5.5 at $4.00, the cheapest model is Claude Sonnet 5.5 at $2.00, 2x less.
Figure 2.
Each point is a model on sale, placed by its index score and its input price. A filled point has a measured score and a ring has an estimated one. Green points are open-weight models, whose weights can be downloaded, and blue points are models sold only through an API. Try: drag the date from 2023 to today to see how the dashed line has moved, and hover any point to see which model it is and where its score and price come from.
Start with the cheapest model that meets your capability bar, and move up only if it fails on your tasks.
The best open-weight model on sale, GLM-5.3, scores 44.8 against 57.6 for the most capable model, and costs $1.40.
This changes if you need low latency, a specific tool-use format, data residency or an enterprise contract.
Limits: Open-weight models do not share one licence. DeepSeek ships under MIT, while the largest recent releases from Qwen and MiniMax carry their own licences with conditions on commercial use. A downloadable model also has to be served by someone, and hosts price it differently. Llama 3.3 70B costs $1.04 per million tokens on Together, and the cheapest of the providers behind OpenRouter lists it at $0.10 in and $0.32 out.
A model with a lower input price can cost more per request
You pay per request, and a request costs input tokens × input price + output tokens × output price. Across the models on sale, the output price runs from 1x to 8.3x the input price, with a median of 4.2x. Grok 4.7 and GPT-5.6 Terra both charge $2.00 per million input tokens, and a request with 8,000 input tokens and 4,000 output tokens costs 1.6x as much on GPT-5.6 Terra.
A reasoning model writes out its working before it answers, and that working is billed as output. OpenAI's guide says that reasoning tokens "still occupy space in the model's context window and are billed as output tokens". Across the 24 models whose token use Artificial Analysis charts, reasoning is a median 67% of output tokens, from 45% to 78%, and those models write a median of 62,200 output tokens per task on the index.
Past 1,500 output tokens the cheapest of the four models in the figure is Grok 4.7, not GPT-5.1.GPT-5.1 charges $1.25 per million input tokens and Grok 4.7 charges $2.00.
Figure 3.
Each curve is one model. A curve is flat on the left, where the input cost dominates, and rises on the right, where the output cost does. Try: set the two sliders to the input and output token counts of your own requests, and hover a curve for that model's prices.
Compare models by cost per request, using token counts you have measured.
Run your own workload through each candidate model, log the billed input and output tokens, then multiply by each model's prices.
A reasoning effort setting, which tells a model how much to reason, changes how many output tokens a request uses, so it changes the bill.
This changes if your inputs are long and your outputs short. Retrieval and classification sit on the flat part of every curve, where the input price decides the cost.
Prompt caching cuts the cost of a 20-turn agent run by 63%
An agent is a program that calls a model in a loop, and it resends its whole context on every turn: the system prompt, the tool schemas and every turn before it. Input tokens therefore grow with the square of the number of turns while output grows in a straight line. In the run below, which is calculated from the list prices of Claude Sonnet 5.5, input is 78% of the bill by turn 20.
With prompt caching, the provider keeps the start of the prompt that it has already processed (the prefix) and charges a fraction of the input price to read it again, 10% of it on Claude Sonnet 5.5.
By turn 20 the cached run costs $0.24 against $0.64 without caching, a saving of 63%.
Figure 4.
The solid part of each line is what the run has cost so far, and the faint part is where it is heading. Caching costs less from turn 2 on. Try: drag the slider to turn 60 to see the gap at the end of the run, or hover the lines for the cost at any turn.
Keep the prompt prefix stable so it caches, and budget at the cached rate.
Put everything static first (system prompt, tool schemas, examples), and never interpolate a timestamp, a request id or a shuffled tool order into it. One changed byte near the front invalidates everything after it.
An estimate made from the list input price puts a 60-turn run at 4.5x its cached cost.
This changes if your loop is slow or bursty. Cache entries expire (five minutes by default on Claude, with a one-hour option that costs more to write), so a run that idles pays full price again.
Limits: The run assumes that every turn reads from the cache and that the context only grows. An agent that prunes or summarises its context changes the prefix, so its later turns read less from the cache. Writing to the cache is priced differently by each provider: Anthropic charges 1.25x the input price for a five-minute entry, OpenAI charges 1.25x on GPT-5.6 and later and nothing on earlier models, and Google bills for the time that a cache you create explicitly is stored.
Model
Provider
Class
Released
Input /1M
Output /1M
Since launch
No model matches that search.
Scope and conventions
The dataset holds 198 models from 13 providers and 270 price points, as of September 28, 2026. Figure 1 draws the 192 that were released in 2023 or later and have a price above zero. Prices are list prices on the standard tier: the peak rate where a provider prices by time of day, the short-prompt rate where it prices by context length, and an introductory or promotional price while it applied (seven of the price points). Figures 1 to 3 leave out two lower rates: the batch rate, for requests that can wait, and the cached rate. The full table gives the cached rate.
The two lines of Figure 1 and the points of Figure 2 cover the models that have both a price and a score. Of the models on sale, twelve have a price and no score, from Qwen3.7-Flash at $0.03 to GPT-5.4 Pro and GPT-5.5 Pro at $30.00, and two are listed as free (GLM-4.5-Flash and GLM-4.7-Flash).
The lower line of Figure 1 reached $0.05 with GPT-5 nano. On June 11, 2026, OpenAI gave notice that it will shut down the model's only dated version on December 11, 2026. A deprecated model does not count as on sale, so the line rose on that day. Counting GPT-5 nano would make the fall 600x.
Figure 1 sorts models into three classes by name. A model is small if the provider's name for it contains the word mini, nano, lite, flash, haiku, small, micro or instant, or a size of 8B or less, or if it is OpenAI's Luna. It is large if the name contains the word pro, opus, max, large, premier, ultra, fable or mythos, or the size 405B, or if it is an o-series, GPT-4.5, Astra or Sol model. Every other model is mid.
All scores on the Artificial Analysis Intelligence Index v4.3.2 were read on one day, September 28, 2026, because the index is revised. Between readings on September 10, 2026 (v4.3) and September 28, 2026 (v4.3.2), 134 of 633 entries changed, where an entry is one model at one reasoning setting. Scores cover 147 of the 198 models, and 88 of those are estimates. Each model has one score, the highest that Artificial Analysis measured for it at any reasoning effort, or its estimate where nothing was measured. A model that has had several dated versions under one name has the score of the version served on each date.
The seven models from 2023 that have a score sit between 5.0 and 7.0 on the index, against 57.6 at the top, and all of those scores are estimates. On GPQA Diamond, 198 science questions that Artificial Analysis has run on models from 2023 and from 2026, GPT-4 scored 35%, and the 21 small models on sale that have been run on it score 36% to 95%.