Skip to main content

Release 2026.10.27

API changes​

Responses API​

POST /v1/responses serves OpenAI's Responses API on every chat model, with the same keys, prices and limits as chat completions. It covers:

  • text and image input;
  • function tools;
  • structured outputs (text.format);
  • reasoning;
  • streaming with Responses events.

The OpenAI SDKs' client.responses.create, client.responses.stream and client.responses.parse work as they are.

The API is stateless, since Serenity Edge stores no prompts or answers. Send the whole conversation in input on every request. store: true, previous_response_id, conversation and background are refused with HTTP 400, and responses cannot be retrieved. See the Responses API guide.

Decision request limits​

POST /v1/decisions and the Jev-compatible POST /v1/systemone accept at most 256 questions per request; POST /v1/decisions also takes at most 255 options per choice question and 10 levels per score question. A larger request is refused, with HTTP 400 (422 on POST /v1/systemone), a message that says which limit was hit, and the parameter at fault in param. See Current limitations.

Clearer validation errors on decisions​

On orion-compass, orion-compass-flash and d3, the error for a malformed decision request now names the field and how to fix it, and the three models refuse the same requests on POST /v1/systemone with the same messages. For example, a choice given as a string gets "each choice must be an object with a 'value' (string or boolean)", and a question whose text is in question instead of instructions is told so. param points at the exact element, such as questions[0].choices[0]. See Errors.


Documentation​

Release 2026.10.14

API changes​

Decision model d3​

d3 (alias vllm-sr/d3) joins the decision models in public preview. It is vLLM Semantic Router's open-weight 27B decision model (Apache-2.0) and first on the text board of the Jev Decision Index. It costs $0.042 per million input tokens, and output is not billed. It reads text only, has a context of 24,576 tokens, and is served on both POST /v1/decisions and POST /v1/systemone.

Embeddings​

New POST /v1/embeddings takes OpenAI's embeddings request, so client.embeddings.create in the OpenAI SDKs works as it is: vectors of 1024 dimensions for up to 512 inputs per request, billed on usage.prompt_tokens. Use a Serenity Edge API key. See Embeddings, rerank and redaction.

Rerank​

New POST /v1/rerank scores up to 1000 documents against a query, in the format of the rerank APIs of Cohere and Jina. Its tokens are estimated as ceil((characters(query) + characters(document)) / 4) per document, with characters counted as Unicode code points, and returned in usage. See Rerank.

Redaction​

New POST /v1/redact finds and masks personal data in up to 256 texts (192 KiB per request) in the privacy filter's own format, with model optional; the text sent is processed in memory only. Its tokens are estimated as ceil(characters / 4) over all the inputs. See Redaction.

Access and errors on the new routes​

The three routes take Serenity Edge API keys; any other key gets HTTP 403 permission_denied with a message that names the route's API. A model sent to a route that does not serve it, or a request the model cannot process (The model server rejected the request. Check the input sizes and parameters.), gets HTTP 400. See Errors and rate limits.


Models​

Embeddings, rerank and redaction models​

Three models join the OpenAI-style list (GET /v1/models?format=openai), each with its route in endpoints and billed on input tokens only:

ModelAliasRoutePrice per million input tokens
qwen3-embedding-0.6bqwen/qwen3-embedding-0.6bPOST /v1/embeddings$0.008
qwen3-reranker-0.6bqwen/qwen3-reranker-0.6bPOST /v1/rerank$0.008
privacy-filteropenai/privacy-filterPOST /v1/redact$0.009

Current prices are listed live on Pricing and limits.

paddleocr-vl-1.6 for document pages​

paddleocr-vl-1.6 (aliases baidu/paddleocr-vl-1.6 and paddlepaddle/paddleocr-vl-1.6) joins the chat models: send one page image with a task prompt (OCR:, Table Recognition:, Formula Recognition: or Chart Recognition:) and it returns the page's text, table, formula or chart data, at $0.016 per million prompt tokens and $0.016 per million completion tokens (a full page is about 1,250 prompt tokens). It reads one image per request, uses temperature 0 unless you send one, has a context of 16,384 tokens and at most 8,192 output tokens, and does not reason, call tools or produce structured outputs. It is available on the API now, and ready for the OpenRouter onboarding, which is not live yet; see Document OCR.


Documentation​

Release 2026.10.13

API changes​

Decisions, in preview​

New POST /v1/decisions (the format of OpenAI's Decisions API) and POST /v1/systemone (compatible with Jev's format) return typed, calibrated answers (probabilities, choices and scores) instead of generated text. Use a Serenity Edge API key. See Decisions.

Image input on the chat models​

POST /v1/chat/completions accepts up to 4 images per request as base64 data URLs in image_url parts, billed as input tokens. Image URLs are refused with HTTP 400. See Models.

Server errors are HTTP 500​

Errors with code upstream_error and timeout now answer HTTP 500 instead of 502 and 504, with the same JSON body and code, so the error object always reaches you. Handle them as before: retry with backoff. See Errors and rate limits.


Models​

Orion Compass decision models​

orion-compass and orion-compass-flash join the catalogue in preview, billed per input token only; output is free. Prices are listed live on Pricing and limits.


Documentation​

  • New guide: Decisions.
  • Models: image input.
  • Errors and rate limits: 422 on /v1/systemone, 403 for keys that cannot use decisions, and 500 for upstream_error and timeout.
  • API reference: the two decision endpoints, image parts in user messages, and the error responses.

Release 2026.09.22

First publication of the Serenity Edge documentation.

Documentation​