What's new

Usage exports match filtered dashboard records

Fixed usage queries and exports for long date ranges and filtered views so exported records match those shown in the dashboard.

Image edits keep the order of input images

Image edits with several input images use them in the order you sent them. If a request runs past your client timeout, you can still get the result from the task record.

Billing shows small charges and refunds

Small charges and refunds are no longer rounded to zero in billing views, so you can see per-request costs below one cent.

Transcription and synchronous speech responses match the docs

Fixed response formats for transcription and synchronous speech endpoints so returned fields and structures match the documentation.

Responses are marked correctly when they reach the length limit

Responses that reach the length limit now carry the correct truncation marker. The interface indicates that the output is incomplete.

Task webhooks retry after delivery failures

Task completion webhooks are retried after a failed delivery, so notifications do not rely on a single attempt.

Video requests without a published price are rejected

A video request is rejected up front when its size and mode combination has no published rate, avoiding acceptance at an unclear price. Reference-to-video uses the same published rates as the rest of the video catalog.

Check the size, mode, and price before submitting a video request.

Text-only requests to vision models work normally

A request with only text, sent to a model that also accepts images, now returns a normal text reply instead of an error.

Short rate limits are retried automatically

When a request hits a short rate limit and a retry is safe, TokenLab retries it for you, so fewer requests fail.

One error format for all models

Errors from every model use the same response format and stable error types, so you can handle failures in one place.

Oversized image requests are now rejected early

Image upload requests that exceed the supported size are rejected before processing. Customers receive a clear error immediately, and the request does not consume usage.