reference

Errors and limits

Errors use the OpenAI envelope, so existing client error handling works:

{
  "error": {
    "message": "Your credit balance is empty. Add credits at ...",
    "type": "insufficient_quota",
    "param": null,
    "code": "insufficient_credit"
  }
}

Codes

status	code	meaning	retry?
400	`invalid_json`	The body was not valid JSON.	No
400	`missing_text`	`/v1/restyle` was called without a non-empty `text`.	No
400	`missing_question`	`/v1/fanout` was called without a non-empty `question`.	No
400	`input_too_large`	Over the input bound for the model you named: 6,000 characters for a restyle, which you send a section at a time, or 2000 for a fan-out question. The message names the length you sent.	No
400	`input_too_long`	A single block exceeded the context window. Add a paragraph break.	No
400	`wrong_endpoint_for_model`	The model named does not do the work this endpoint does — a rewrite model at `/v1/fanout`. Served badly it would return a plausible-looking answer, so it is refused.	No
401	`invalid_api_key`	Missing, malformed, unknown or revoked key.	No
402	`insufficient_credit`	The account balance is empty. Add credit, or turn on auto top-up.	After topping up
402	`key_spend_limit_reached`	This key hit its lifetime cap. Other keys are unaffected.	After raising it
403	`model_not_permitted`	The key’s allowlist does not include this model.	No
404	`model_not_found`	No such model id.	No
404	`model_not_available`	The model is in the catalogue but not serving yet.	No
502	`empty_generation`	A fan-out produced nothing usable for that question, so there is no list to return.	Once, then rephrase
502	`upstream_error`	The model returned an error.	Yes, with backoff
502	`upstream_unreachable`	The model could not be reached.	Yes, with backoff
503	`no_capacity`	No capacity is available. A problem on our side, not with your request.	Yes, with backoff
504	`timeout`	The model took too long.	Yes

In short: retry 5xx with exponential backoff and jitter — with the exception of empty_generation, where the model ran and simply produced nothing usable, so one retry is worth it and a loop is not. Do not retry 4xx, nothing about the request will have changed.

Failed requests are not charged

A request that fails before the model generates anything costs nothing. It is still recorded, at zero cost, so a failing integration is visible on your usage dashboard instead of only in our logs.

Limits

limit	value
Request body, restyle	6,000 characters per request
Single block, restyle	~1,150 words
Question, fan-out	2000 characters
Queries returned, fan-out	Up to 6, after near-duplicates are dropped
Context window	4,096 tokens, prompt and completion together
Request timeout	180 seconds per block
Playground, signed out	10 runs a day per IP, 2,000 characters each

Per-key concurrency limits apply so that one caller cannot occupy the whole fleet. If you are planning something high-volume, email hello@meragpt.com first and we will make sure the capacity is there.