ModelShelf
Every AI model's price, and how long it lasts.
Rebuilt every day from the full git history of LiteLLM's model price file. Of the 5,257 models added since 6 September 2023, these are the estimated shares whose first change was each of the following.
Not financial or procurement advice. The numbers describe edits to a public catalogue, not what any model costs you. Scope and disclaimer.
The maker and the grouped names come from naming rules (for
example, eu.anthropic.claude-3-5-sonnet-20241022-v2:0 and
snowflake/claude-3-5-sonnet are both grouped as anthropic /
claude-3.5-sonnet). The host is LiteLLM's own provider field.
What would it cost?
Paste a prompt, or type how many tokens it has. Your text stays in this page: it is not sent anywhere or saved.
Add images, documents, audio or video to the prompt
Only models that can take these in are shown. When LiteLLM has no per-image or per-second price for a model, an image counts as about 1,000 tokens, a PDF page about 1,500, a second of audio about 32 and a second of video about 300.
The model must be able to
These are estimates. Prompt tokens are counted at about 4 characters per token, a common rule of thumb for English; each model counts its own way, often 10 to 30% differently. You cannot know the answer length in advance, so pick the closest size. Thinking models also bill their hidden reasoning as answer tokens, often several times the visible answer. "How many times" is how often you send this prompt, for example 1,000 customer questions a day; in a chat, each message also re-sends the conversation so far. Caching, batch discounts and long prompt prices are left out.
Best value for a task
Quality comes from the Arena leaderboard, where people compare two anonymous answers to the same prompt and vote for the better one. Each rating has a 95% range. Best value is the cheapest model whose range overlaps the top model's, so the votes cannot tell it apart from the best.
How often is the automatic guess right?
Changed in the last 7 days
Every price change, retirement date, removal and addition in the catalogue in the 7 days before the latest data. A single clean up can remove hundreds of rows at once. Tags mark bulk clean ups (50 or more models removed on the same day), unit fixes (a price made 100 times larger or smaller) and reverted changes (undone within 30 days).
Choose makers, models, versions and hosts
Leave a list on All to include everything. Models and versions are split from the grouped names by naming rules, which can be wrong.
Most repriced
Counts changes from one price to another. A price appearing for the first time is not counted, and neither are unit fixes or changes undone within 30 days.
Hosts ranked by chance of a price change within 1 year
Only hosts with at least 100 models still followed at one year are ranked; smaller hosts would give unreliable numbers.
First exit, by time since the model was added
Removals include bulk clean ups of the catalogue, not only retired models; the last table below shows the numbers without them. Hover over a bar to see its value.
The numbers, with 95% confidence intervals
| Within | Price change | Retirement announced | Removed | None of these | Still followed |
|---|---|---|---|---|---|
| 30 days | 5.9%5.3 to 6.6 | 3.5%3.0 to 4.0 | 10.1%9.3 to 10.9 | 80.5%79.4 to 81.6 | 3,697 |
| 90 days | 8.6%7.8 to 9.4 | 5.2%4.5 to 5.8 | 13.2%12.3 to 14.1 | 73.1%71.8 to 74.3 | 2,717 |
| 6 months | 10.9%10.0 to 11.8 | 7.7%6.9 to 8.5 | 14.1%13.1 to 15.1 | 67.4%66.0 to 68.7 | 2,336 |
| 1 year | 14.7%13.6 to 15.8 | 16.4%15.2 to 17.6 | 16.2%15.2 to 17.4 | 52.7%50.9 to 54.3 | 1,214 |
| 2 years | 18.3%17.0 to 19.6 | 29.5%27.6 to 31.4 | 23.3%21.6 to 24.8 | 28.9%26.9 to 31.2 | 379 |
Aalen-Johansen estimates of competing risks with 1,000 bootstrap resamples. Each row adds up to 100% because the exits compete: once a model is removed, its price can no longer change. Download as CSV.
Does the definition matter?
| What counts as a price change | Within 1 year |
|---|---|
| Any price field (main result) | 14.7%13.6 to 15.8 |
| Input and output token prices only | 12.5%11.6 to 13.5 |
| Ignoring changes in the first week | 12.1%11.0 to 13.1 |
| Leaving out corrections | 12.4%11.3 to 13.5 |
Every change, or real changes only?
Some edits fix the catalogue rather than change a model. Real changes only leaves out unit fixes (a price made 100 times larger or smaller) and price changes undone within 30 days, and treats a model removed in a bulk clean up (50 or more models removed on the same day) as still followed until that day rather than removed.
| Within 1 year | Price change | Retirement announced | Removed | None of these |
|---|---|---|---|---|
| Every change (main result) | 14.7%13.6 to 15.8 | 16.4%15.2 to 17.6 | 16.2%15.2 to 17.4 | 52.7%50.9 to 54.3 |
| Real changes only | 12.4%11.3 to 13.5 | 18.7%17.3 to 20.0 | 5.7%5.0 to 6.5 | 63.3%61.5 to 64.9 |
Scope and disclaimer
- Not financial, procurement, legal or tax advice. Nothing here says which model or provider to use, or what any model costs you.
- Catalogue edits, not provider prices. Every number describes edits to LiteLLM's
model_prices_and_context_window.json: when an entry was added, when its listed price changed, when a retirement date was set and when it was removed. A provider can change its real price before, after or without such an edit. - Past patterns in the catalogue say nothing certain about future changes. The confidence intervals describe sampling uncertainty only.
- Independent: the data comes from the MIT licensed BerriAI/litellm repository. Not affiliated with or endorsed by BerriAI or LiteLLM.
- Provided as is, without warranty of any kind.