OpenAI Responses
/v1/responses compatible endpoint for Codex-style tools and multi-turn previous_response_id chains.
Endpoint
POST/v1/responses
| Field | Type | Description | |
|---|---|---|---|
| model | string | required | Model name. |
| input | string | array | required | Text or structured input items. |
| instructions | string | optional | System instructions. |
| previous_response_id | string | optional | Continue from the previous response. |
| tools | array | optional | Tool definitions, passed through upstream. |
| stream | boolean | optional | SSE event stream: response.created … response.output_text.delta … response.completed. |
curl https://www.link42.ai/v1/responses \
-H "Authorization: Bearer $LINK42_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "deepseek-v4-flash", "input": "Write a Python function that checks whether a number is prime"}'- Codex CLI / OpenAI Agents SDK only need base_url pointed at https://www.link42.ai/v1.
Model reference
Responses is the native shape for both models below: the private protocol submits them with a top-level input[], and the standard surface uses the /v1/responses endpoint on this page.
| Model id | Name | Upstream | Billing | Fields that differ | Limits and gotchas |
|---|---|---|---|---|---|
| seed-1-6-250915 | Seed-1.6 | BytePlus Ark | Per token | input[].content accepts input_text and input_image; besides message items, output[] also carries reasoning items (summary[].summary_text); usage reports input_tokens / output_tokens with input_tokens_details.cached_tokens and output_tokens_details.reasoning_tokens | Above 128,000 input tokens in one request the whole call is priced at double rate; image input counts inside input_tokens at the text input price |
| deepseek-v3-2-251201 | DeepSeek-3.2 | BytePlus Ark | Per token | tools accepts [{"type": "web_search", "max_keyword": 3}]; when a search runs, output[] gains web_search_call items (with action.query) and usage gains tool_usage.web_search and tool_usage_details.web_search | The long-context threshold is 32,000 tokens, lower than the Seed family; above it the whole call is priced at double rate. Tool calls themselves are not billed, only tokens |
- With stream=true a standard_v1 key gets real SSE relayed frame by frame, while a legacy_v1 key gets the whole exchange buffered into one body — the old system's pseudo-streaming.