The same intelligence,
two years apart.
Each square is what a million tokens of GPT-4-class capability cost that quarter, from the cheapest model that cleared the bar. Area is price.
Comparing this year's flagship with last year's measures two things at once, because models get cheaper and better. So this holds capability fixed and asks only: what did the cheapest model that clears the bar cost?
The frontier cost curve
Each line answers one question: on this date, what was the cheapest model that could clear this bar? Pick the benchmark the bar is measured on — science, coding or competition math. Prices are blended 75% input / 25% output, log scale — a straight line here is a constant multiple per year.
Every model we have a price history for
Each row is one model's full recorded price history. The sparkline is its blended price over time; the drop column compares its first recorded price to its latest.
| Trend |
|---|
What your app would have cost
The frontier curve in the only units that matter to a founder: your monthly bill per user, if you'd shipped the same product at each point in the past.
Take the data
Everything on this page is static JSON and CSV. No API key, no rate limit — fetch it, fork it, correct it.
| File | What's in it |
|---|---|
| data/prices.csv | One row per price change: model, provider, date, input, output, cache read/write ($/1M tokens) |
| data/models.json | Full registry — price history with per-event provenance, aliases, context window, benchmark scores |
| data/frontier.json | The frontier curve, both provider scopes, all capability tiers |
| data/index.json | Summary counts and coverage |
| pipeline/ | Every script used to collect, clean and validate the data |
Every price on this site links to its own source — the LiteLLM commit or the archived vendor page it was read from. Open any model page to check a figure yourself. The repository is at github.com/hiteshbandhu/token-price-history.