The provider index shows what open models cost through every provider that public sources list. This page covers where each number comes from and the rules for every row.
Pareto Inference runs this page and is one of the providers listed. Rows are sorted by price, and Pareto's rows are labeled and placed by price like any other. The methodology says what is included and what is left out. Spot a wrong number? Email us.
Sources
- OpenRouter. Its public list of each model's providers: price, cached-input price, context, maximum output and uptime over the last day. Listed prices already include OpenRouter's promo discounts; we mark the discount.
- Vercel AI Gateway. Its public list of each model's providers: global price (not regional prices), peak pricing, uptime over the last day, and the median output speed and time to first token over the last hour of its live traffic.
- models.dev. The open model catalog, used only for providers that neither gateway lists, including some labs' own APIs.
- Pareto. Pareto's own self-serve prices, the same as on Pricing.
- Feeds, not scraping. Only the sources' public JSON feeds, nothing behind a login. Every row names its source; hover a source to see when it was checked.
Rules
- Unit. USD per 1M tokens for input, output and cached input. A dash means the source lists no price, not a free one.
- Peak and off-peak. Time-of-day prices show both prices and their hours in UTC. Public holiday exceptions are named but not applied to the price shown now.
- One row per offer. A provider's offers at the same prices on several sources share a row that names each source. Different prices, tiers or precisions get their own rows.
- models.dev rows. One per provider: its cheapest standard variant of the model. Renamed variants (faster, free, regional, fine-tuned), free tiers and subscription plans are left out.
- Not included. Taxes, fees for buying credits, and account-specific or volume discounts.
Sorting
- Order. By the agent blend at the moment the page is served: the price of 1M tokens that are 5% uncached input, 94% cached input and 1% output. Ties sort by the 3:1 blend.
- Why that mix. It is Pareto's own traffic over the 28 days to November 6, 2026: 44 trillion tokens from 139 accounts, nearly all coding agents, were 4.1% uncached input, 95.3% cached input and 0.6% output. Each model, the traffic without its largest account, and the median account all came out between 92% and 97% cached input. At these shares the cached-input price is most of an agent's bill.
- What it favors. Providers with cheap cached input rank higher, Pareto included. A listed cached-input price doesn't promise the same cache hit rate at every provider, and a provider that lists none is counted at its full input price.
- The 3:1 blend. Three input tokens to one output token with the cache ignored, the common comparison for chat traffic, is shown under every agent blend and in the downloads.
- Cheapest right now. The lowest price for each blend and token type at the same moment, peak hours applied.
- Pareto's rows. Labeled "Runs this page", placed by price like any other row, and shown with whatever the gateways measure, weak numbers included.
- Speed and uptime. Only the gateways' own measurements, so direct rows, Pareto's included, show none.
Freshness
- Refresh. Each source is read at most once every hour. If a read fails, the last good copy stays; a source that has never answered is listed as unavailable.
- Reuse. Every model's table downloads as CSV or JSON, with each row's sources and check time. Cite the model's page.
- Corrections. If a number is wrong, email us with the page and the row. A provider's own price changes at its source and shows here within the hour.