Model tracker: what the leading models cost, and what your workload would cost
Model prices change every few weeks, and the cheapest model per token is often not the cheapest per month. This tracker lists the current list prices of the leading models with the page each was read from, and turns a workload you describe into a monthly bill for each of them.
Level BeginnerAbout 8 minRead first: KV cache, Quantisation1 challenge in the Lab
Interactive simulation. It needs JavaScript; the text below explains what it shows.
What you are looking at
The bars show the monthly bill of the chosen workload on every model, cheapest first. A workload is three numbers: input tokens a month, output tokens a month, and the share of input that hits the provider's prompt cache and is billed at the cached rate. The table below lists the standard list price per million tokens for each model, with a link to the page it was read from and any notes, such as a higher rate for very long prompts or a price change that is already announced.
Prices are read again every morning by the same automated job that writes this site's daily articles; the date above the bars says when.
Things to try
- Start with the support chatbot, then switch to the coding agent: the order of the models changes, because the agent's bill is dominated by input.
- On the coding agent, move cache hits from 85% to 0%: the bill of the expensive models multiplies, the cheap ones barely move.
- Enter your own numbers: a month of logs usually gives input and output tokens directly.
How it works in real systems
Real bills depend on more than list prices: reasoning models spend hidden thinking tokens, tokenizers turn the same text into different numbers of tokens, batch APIs halve the price for work that can wait a day, and long prompts can move a request into a higher price tier. Teams usually measure cost per completed task on their own traffic, then route easy requests to cheaper models. The tracker gives the starting point for that measurement.
Key formulas
$$\text{cost}=\frac{I\,(1-h)\,p_{\text{in}}+I\,h\,p_{\text{cached}}+O\,p_{\text{out}}}{10^{6}}$$
Monthly cost from input tokens I, output tokens O, cache-hit share h and prices per million tokens.
Common mistakes
- Comparing models by price per token alone: models that think longer or tokenize your text into more tokens cost more per task.
- Forgetting the context-length tiers: several providers charge more above a prompt-size threshold.
- Ignoring caching: for agents that resend the same long context, the cached-input price matters more than the input price.
Real-world numbers
- The list on this page was read on 20 November 2026 from the providers' pricing pages and is checked again each morning.
- Gemini 4 Argon and Carbon prices double on 1 December 2026; the tracker switches automatically on that date.
Further reading
Code lab
The same calculation in Python, for your own workload and price list.
Step 1 · A price list and a workload
prices = { # per million tokens: input, cached input, output
"Claude Sonnet 5.5": (2.0, 0.10, 10.0),
"GPT-6.2 Sol": (2.0, 0.10, 10.0),
"Gemini 4 Hydrogen": (1.5, 0.075, 7.5),
}
workload = dict(input=900e6, output=25e6, cache_hit=0.85)
Step 2 · Monthly cost per model
def monthly(p, w):
i, c, o = p
return (w["input"] * (1 - w["cache_hit"]) * i + w["input"] * w["cache_hit"] * c + w["output"] * o) / 1e6
for name, p in sorted(prices.items(), key=lambda kv: monthly(kv[1], workload)):
print(f"{name:20s} ${monthly(p, workload):,.0f} a month")