Limits

These are the values on proxium.tech today. A limit of 0 means no limit.

The limits you meet first

Limit	Value
Request body	32 MiB
`x-proxium-timeout-ms`	At most 120000 ms
Models in one plan	4
Retry-After that Proxium waits	At most 30 s
App limits	None until you set them

Limit	Value	At the limit
Request body	32 MiB	`413`, plain-text body
Vendor answer	64 MiB	The call fails. Proxium tries the next model
Stored copy of an answer	32,768 characters	Proxium cuts the stored copy. Your app gets the full answer
Model id of a media call	200 characters	`400 bad_model`

A media model id can hold only A-Z, a-z, 0-9, ., _, - and :.

Limit	Value	At the limit
`x-proxium-timeout-ms`	At most 120000 ms	A larger value is lowered to the maximum
All calls of one request, with the header	The header value	`503 upstream_failed`
All calls of one request, without the header	No limit
One call, not streamed	120 s	Timeout. Proxium tries the next model
Connect to a vendor	10 s	The call fails. Proxium tries the next model
Silence in a stream	300 s	Proxium ends the stream

Proxium ignores an x-proxium-timeout-ms of 0, a negative value, or text.
When the header time ends, the message of the 503 starts with client deadline exceeded.
For a stream, the header time ends at the first piece. After that, only the silence limit applies. A stream has no limit on its total time.
The silence limit also runs while the vendor prepares its first piece.

Limit	Value	At the limit
Models in one plan	4 by default, 1 to 4	`503 upstream_failed`
Extra calls to one model	1 by default, 0 to 3	Proxium tries the next model
First wait before an extra call	200 ms by default, 0 to 5000 ms. It doubles for each next call	—
Retry-After that Proxium waits	30000 ms at most by default, 0 to 30000 ms	Proxium waits, then calls again
Failures that open a breaker	5 in 10 minutes	Proxium skips the vendor
Pause after normal failures	30 s	One trial call
Pause after a rejected key or used quota	30 minutes	One trial call
Trial calls that close a breaker	2 good answers	The vendor is used again
Failed calls in a row that retire a model	3 by default, 1 to 100	The project's calls skip the model
Time a retired model stays retired	60 minutes by default, 1 to 10080	The model is back
Time a vendor stays retired after "no credits" or "key rejected" on your own key	60 minutes by default, 1 to 10080	The vendor is back, and its first call is the test call
Failed calls in a row across its models that retire a vendor	5 by default, 1 to 100	The project's calls skip every model of the vendor
Time a vendor stays retired after failed calls	60 minutes by default, 1 to 10080	The vendor is back

An extra call to a model happens only after a 5xx, 408 or "warming" answer. After a 429, by default, the next call goes to a model at another vendor of the plan. With no other vendor, the same model gets one more call.
A 429 never opens a breaker, and never counts toward a retirement.
Your project sets each "by default" value of this table in its failover policy, on Routing › When a call fails. See Set your failover policy.
Proxium does not skip a fallback model for its price. A fallback can cost more than your first model.

Limit	Value	At the limit
App: Calls per hour	You set it. Default `0`	`429 source_capped`
App: Dollars per day	You set it. Default `0`	`429 source_capped`
Key: calls per hour	You set it on Keys. Default `0`	`429 tenant_capped`
Key: requests and tokens per minute	You set them on Keys. Default `0`	`429 tenant_capped`
Key: dollars per day	You set it on Keys. Default `0`	`429 tenant_capped`
Project: monthly cap	None during the open beta

Each 429 carries retry-after: 60.
Calls per hour: the last 60 minutes, rolling.
Dollars per day: the last 24 hours, rolling, in hourly buckets.
Requests and tokens per minute: the last 60 seconds, rolling.
You set the app limits on Overview › Per-app burn, and the key limits on Keys. See Set budgets and limits.

Limit	Value	At the limit
Key name, on Keys	60 characters	The name is dropped. The key has no name
Endpoint name	60 characters	`400 bad_name`
Vendor name	40 characters	`400 bad_name`
Models of one vendor	200	`400 too_many_models`

A vendor name holds only a-z, 0-9 and -, with no - at the start or the end.

Change	Applies
Routing, app limits, the cache price	At once on one server, within 30 s on all
The spend that the monthly cap reads	Every 30 s
The model list of each vendor	Every 21600 s