Zero Completion Insurance | OpenDunes Docs

Not yet available. Zero completion insurance is on the roadmap.

A model call can fail to produce meaningful output for reasons entirely outside your control — the upstream provider returns an empty response, cuts off mid-stream, or returns an error. Zero completion insurance means you are not billed for those cases. If the model doesn't deliver, you don't pay.

What counts as no usable content#

Zero completion insurance applies when:

The response body contains no choices, or choices is an empty array.
Every choice has finish_reason: "error" or an empty message.content.
A streaming response ends with [DONE] but no content deltas were emitted.
The request results in a 503 temporary_unavailable or 502 upstream_error/upstream_timeout error from OpenDunes.

What is still charged#

Not every failure is covered. You are still charged when:

The model returns content but you discard it (that's a product decision, not a model failure).
The request fails due to your own error — 400 invalid_request, 401 unauthenticated, 402 insufficient_credits.
The model returns partial content in a streaming response before stopping mid-stream due to max_tokens being reached — partial output is usable output.

How the credit hold works#

When you send a request, OpenDunes places a temporary hold on the estimated cost (based on input tokens). The hold is released after the response is received:

Content returned → hold is converted to a final charge for actual input + output tokens.
No usable content → hold is released in full. Your balance is not reduced.

The final charge (or release) appears in your activity log within seconds of the response completing.

Checking coverage in the activity log#

Every request in the activity log shows a billed field. Requests covered by zero completion insurance appear with billed: 0 and a zero_completion flag.

Provider reliability#

Model reliability varies by provider and model. You can review uptime and error rates per model in the model catalog under the Reliability tab — this helps you choose models with the best track record for your use case.

Note. Zero completion insurance does not substitute for retry logic. For production workloads, implement exponential backoff on 503 and 504 errors and consider configuring model fallbacks to automatically reroute to an alternative model when the primary is unavailable.

What counts as no usable content#

Zero completion insurance applies when:

The response body contains no choices, or choices is an empty array.

Every choice has finish_reason: "error" or an empty message.content.

A streaming response ends with [DONE] but no content deltas were emitted.

The request results in a 503 temporary_unavailable or 502 upstream_error/upstream_timeout error from OpenDunes.

What is still charged#

Not every failure is covered. You are still charged when:

The model returns content but you discard it (that's a product decision, not a model failure).

The request fails due to your own error — 400 invalid_request, 401 unauthenticated, 402 insufficient_credits.

The model returns partial content in a streaming response before stopping mid-stream due to max_tokens being reached — partial output is usable output.

How the credit hold works#

When you send a request, OpenDunes places a temporary hold on the estimated cost (based on input tokens). The hold is released after the response is received:

Content returned → hold is converted to a final charge for actual input + output tokens.

No usable content → hold is released in full. Your balance is not reduced.

The final charge (or release) appears in your activity log within seconds of the response completing.

Provider reliability#

Note. Zero completion insurance does not substitute for retry logic. For production workloads, implement exponential backoff on 503 and 504 errors and consider configuring model fallbacks to automatically reroute to an alternative model when the primary is unavailable.