Models & Pricing
What the :free suffix means, how usage is priced and why a model sometimes disappears.
The Models page lists every model with context window, per-token prices, supported endpoints and capability filters like tool calling and image input. Each model has its own page with copy-paste code examples. Not every free model supports tool calls or vision, so check the capability badges before wiring one into a coding agent.
Rankings shows which free models actually deliver, based on community test results, and Status tracks live provider health.
New here? Getting a key and sending a first request takes a minute in Quickstart
Free models carry a :free suffix, for example gpt-oss-120b:free
A :free model routes only to free upstream providers and never touches your balance. The same base name without the suffix is the paid version: stable, uncapped and billed per token. Both can exist side by side, so switching from free to paid is a one-string change.
Free models have two layers of limits. The upstream providers cap their free pools, and on top of that we add our own per-user, per-model cap so one heavy user cannot drain a shared pool. Expect 429 responses at peak times, and use a paid model when you need reliability.
The exact caps, headers and retry guidance are documented in Errors & Rate Limits
Every model routes across one or more provider groups, each with its own price; the Group Pricing table on the model page lists them all. High availability earns automatic discounts, and a failover to another group can shift the effective price.
To lock a model to specific groups and their rates, see Group Pinning
Most models bill per token, with separate input and output prices. A few models (mostly image and video) bill a flat price per call instead. What you see on the model page is what you pay: no subscriptions, no hidden fees, your balance simply decreases per request.
The Pricing page has current top-up options; every model page shows live per-token prices.
For models that support prompt caching (Claude and others), repeated prompt prefixes are billed at a reduced cached-input rate, while writing a new cache entry costs slightly more than a normal input token (about 1.25x).
Caching is automatic. Workloads with long stable system prompts (agents, RP presets) profit the most, with no configuration needed.
Requests fail over automatically to the next provider group when one hits its rate limit or goes down. Only when every channel serving a model is out does the model disappear from the catalog.
A model vanishing under load is expected, not an outage: it reappears on its own once a channel recovers, usually within minutes.
Keys that pin provider groups fail over only within their pinned groups, see Group Pinning
To get pinged the moment it comes back, watch it in Notifications