Your ideas.
Our API.

Text models

Select an enabled text model from GET /v1/models. Example IDs illustrate request syntax and do not indicate current availability. Use Responses or Chat Completions for GPT models. Messages is for Claude and compatible models. Vision, tools, JSON Schema and reasoning capabilities depend on the selected model; changing the protocol does not add capabilities.

GPT-5.6 Luna

gpt-5.6-luna accepts text and images and returns text. Use POST /v1/responses with input or POST /v1/chat/completions with messages; /v1/messages is not supported for this model. Images use input_image in Responses or image_url in Chat, with public HTTPS URLs or data URLs.

Use reasoning.effort in Responses or reasoning_effort in Chat: low, medium, high, xhigh, max; the default is medium. Context is 1,050,000 tokens, input up to 922,000, output up to 128,000. Reasoning tokens count toward the output cap and are billed as output tokens.

service_tier is an optional string. Omit it for the standard tier. priority or fast selects Fast: every published RUB rate for this model is multiplied by 2, including input, output, cache reads and cache writes. This rule is for Luna; check other models separately for Fast support.

GLM-5.3 Flash

For glm-5.3-flash, use POST /v1/chat/completions. The model accepts text and image_url inputs with public HTTPS URLs or data URLs and returns text. Context is 1,000,000 tokens, output up to 128,000. Thinking is always enabled; reasoning_effort accepts high or max, default max. low and medium run as high; xhigh runs as max. service_tier is not supported: omit it.

Responses and Messages are available through conversion to Chat Completions with limitations. Responses accepts text or input_image; input_file is not carried over. Messages supports text, tool blocks and base64 images only; URL images are not carried over. Web search and other server-side tools are unsupported in both conversions. reasoning.effort in Responses uses the same levels. Set output_config.effort explicitly in Messages: when omitted, medium is sent, which this model runs as high.

All three methods accept JSON and require Authorization: Bearer, Content-Type: application/json and a unique Idempotency-Key for every new operation. In these examples API_BASE is the base URL shown above, without a trailing /v1, and API_KEY is your key. Create REQUEST_KEY once and save it with the request body. Reuse it for the same operation; create another key for a new request.

REQUEST_KEY="$(python3 -c 'import uuid; print(uuid.uuid4())')"

Text model limits and capabilities

Reviewed 2026-09-28. These are declared technical capabilities, not a promise that every ID is available: an enabled model must appear in GET /v1/models. Context includes input and output together; maximum input and output may not be achievable simultaneously. System instructions, history, tools and reasoning also consume tokens. Not declared means there is no confirmed value; it does not mean zero, unlimited capacity or support for a feature. The API validates the common format; a model may reject requests that exceed its own limit.

The next table separates listed levels, additional accepted values and declared defaults. An absent default does not imply medium. For GPT models with default=none, do not replace an omitted field with an arbitrary effort level. The effort field depends on the protocol, as described below. See Reasoning / thinking for additional GLM/Kimi rules.

Claude Opus 5: thinking can be disabled only at effort high or lower. Claude Fable 5.1: do not force a tool with tool_choice any/tool. Claude Sonnet 5: omit temperature, top_p and top_k; arbitrary overrides of these settings are unsupported. Current limits are unconfirmed for deepseek-v4-flash-vision-exp, deepseek-v4-pro-0813 and gpt-5.3-codex-spark; settings from other IDs do not automatically apply to them.

POST /v1/responses

curl --fail-with-body -sS -N "$API_BASE/v1/responses" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Idempotency-Key: $REQUEST_KEY" \
  -H "Content-Type: application/json" \
  --data '{
  "model": "gpt-5.6-sol",
  "input": "Explain a database transaction in two sentences.",
  "max_output_tokens": 512
}'

HTTP 200: traverse output[] and collect content[] blocks with type=output_text. The array can also contain reasoning and function_call items. Do not assume a fixed text index. status=incomplete means a partial result; inspect incomplete_details.

{
  "id": "8488e6a4-d38a-4f95-a831-752aa12bf5d7",
  "object": "response",
  "status": "completed",
  "output": [
    {
      "type": "message",
      "role": "assistant",
      "content": [
        {
          "type": "output_text",
          "text": "Ready."
        }
      ]
    }
  ],
  "usage": {
    "input_tokens": 12,
    "output_tokens": 3,
    "total_tokens": 15
  }
}

POST /v1/chat/completions

curl --fail-with-body -sS -N "$API_BASE/v1/chat/completions" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Idempotency-Key: $REQUEST_KEY" \
  -H "Content-Type: application/json" \
  --data '{
  "model": "glm-5.3",
  "messages": [
    {
      "role": "user",
      "content": "Explain a database transaction in two sentences."
    }
  ],
  "max_completion_tokens": 2048,
  "reasoning_effort": "low"
}'

HTTP 200: text is in choices[].message.content, function calls in message.tool_calls and the termination reason in finish_reason. length indicates that a cap was reached and text may be truncated.

{
  "id": "8488e6a4-d38a-4f95-a831-752aa12bf5d7",
  "object": "chat.completion",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Ready."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 12,
    "completion_tokens": 3,
    "total_tokens": 15
  }
}

POST /v1/messages

Send anthropic-version: 2023-06-01. You may use x-api-key with your key when Authorization is absent. Use the optional anthropic-beta header only for a confirmed feature of the chosen model.

curl --fail-with-body -sS -N "$API_BASE/v1/messages" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Idempotency-Key: $REQUEST_KEY" \
  -H "Content-Type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  --data '{
  "model": "claude-sonnet-5",
  "max_tokens": 512,
  "messages": [
    {
      "role": "user",
      "content": "Explain a database transaction in two sentences."
    }
  ]
}'

HTTP 200: content[] contains text, tool_use and other model blocks. Read every text block; stop_reason=max_tokens indicates capped output, while tool_use requests a tool. The usage counters are input_tokens and output_tokens.

{
  "id": "8488e6a4-d38a-4f95-a831-752aa12bf5d7",
  "type": "message",
  "role": "assistant",
  "content": [
    {
      "type": "text",
      "text": "Ready."
    }
  ],
  "stop_reason": "end_turn",
  "usage": {
    "input_tokens": 12,
    "output_tokens": 3
  }
}

Tools and JSON Schema

The bodies below use the corresponding endpoints and the same headers. Your application executes functions after checking arguments and permissions. Send the result in a subsequent request with a new Idempotency-Key. Responses uses an input item with type=function_call_output, call_id and string output. Chat uses a role=tool message with tool_call_id and string content after the assistant tool_calls message. Messages uses a user tool_result block with tool_use_id and content after the assistant tool_use block. Preserve call identifiers and conversation history.

responses — tools

{
  "model": "gpt-5.6-sol",
  "input": "Check order A123 using the available tool.",
  "max_output_tokens": 512,
  "tools": [
    {
      "type": "function",
      "name": "get_order",
      "description": "Read an order status.",
      "parameters": {
        "type": "object",
        "properties": {
          "order_id": {
            "type": "string"
          }
        },
        "required": [
          "order_id"
        ],
        "additionalProperties": false
      },
      "strict": true
    }
  ],
  "tool_choice": "auto"
}

chat/completions — tools

{
  "model": "gpt-5.6-sol",
  "messages": [
    {
      "role": "user",
      "content": "Check order A123 using the available tool."
    }
  ],
  "max_completion_tokens": 512,
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_order",
        "description": "Read an order status.",
        "parameters": {
          "type": "object",
          "properties": {
            "order_id": {
              "type": "string"
            }
          },
          "required": [
            "order_id"
          ],
          "additionalProperties": false
        }
      }
    }
  ],
  "tool_choice": "auto"
}

messages — tools

{
  "model": "claude-sonnet-5",
  "max_tokens": 512,
  "messages": [
    {
      "role": "user",
      "content": "Check order A123 using the available tool."
    }
  ],
  "tools": [
    {
      "name": "get_order",
      "description": "Read an order status.",
      "input_schema": {
        "type": "object",
        "properties": {
          "order_id": {
            "type": "string"
          }
        },
        "required": [
          "order_id"
        ]
      }
    }
  ],
  "tool_choice": {
    "type": "auto"
  }
}

For a strict schema, list required properties in required; additionalProperties=false excludes extra keys. Validate returned JSON in your application and handle model refusals and token caps. Acceptance of a field alone does not establish structured-output support.

responses — JSON Schema

{
  "model": "gpt-5.6-sol",
  "input": "Return the word ready in a status field.",
  "max_output_tokens": 512,
  "text": {
    "format": {
      "type": "json_schema",
      "name": "status_result",
      "strict": true,
      "schema": {
        "type": "object",
        "properties": {
          "status": {
            "type": "string"
          }
        },
        "required": [
          "status"
        ],
        "additionalProperties": false
      }
    }
  }
}

chat/completions — JSON Schema

{
  "model": "gpt-5.6-sol",
  "messages": [
    {
      "role": "user",
      "content": "Return the word ready in a status field."
    }
  ],
  "max_completion_tokens": 512,
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "status_result",
      "strict": true,
      "schema": {
        "type": "object",
        "properties": {
          "status": {
            "type": "string"
          }
        },
        "required": [
          "status"
        ],
        "additionalProperties": false
      }
    }
  }
}

Image input for text models

Choose a vision-capable model. Replace https://example.com/reference.jpg with a directly accessible HTTPS image URL. Base64 support is model-specific: Responses input_image.image_url and Chat image_url.url use data:image/jpeg;base64,...; Messages uses a source with type=base64, media_type=image/jpeg and raw base64 data. For Kimi K3, use base64 data URLs. Image size/count limits and file or audio input support are not shared across all text models.

responses — vision

{
  "model": "gpt-5.6-sol",
  "input": [
    {
      "role": "user",
      "content": [
        {
          "type": "input_text",
          "text": "Describe the image."
        },
        {
          "type": "input_image",
          "image_url": "https://example.com/reference.jpg"
        }
      ]
    }
  ],
  "max_output_tokens": 512
}

chat/completions — vision

{
  "model": "gpt-5.6-sol",
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "Describe the image."
        },
        {
          "type": "image_url",
          "image_url": {
            "url": "https://example.com/reference.jpg"
          }
        }
      ]
    }
  ],
  "max_completion_tokens": 512
}

messages — vision

{
  "model": "claude-sonnet-5",
  "max_tokens": 512,
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "Describe the image."
        },
        {
          "type": "image",
          "source": {
            "type": "url",
            "url": "https://example.com/reference.jpg"
          }
        }
      ]
    }
  ]
}

Video input in Qwen3.8-Max

The original Qwen3.8-Max model can understand videos, but the current Neironych route supports text and image input only. The video_url, input_video and video parts return HTTP 422 invalid_request_contract with X-Validation-Error: video_input_unavailable before task creation or charging. Do not silently replace video with still frames. Text requests require an Idempotency-Key of 8 to 160 characters.

Reasoning / thinking

For GLM and Kimi, use reasoning_effort through Chat Completions. The table applies when the corresponding model is available; levels are not universal. GPT uses reasoning.effort in Responses or reasoning_effort in Chat. Claude thinking and output_config depend on generation; omit them unless the configuration is confirmed. HTTP 200 does not prove that an arbitrary unsupported setting affected the result.

curl --fail-with-body -sS -N "$API_BASE/v1/chat/completions" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Idempotency-Key: $REQUEST_KEY" \
  -H "Content-Type: application/json" \
  --data '{
  "model": "kimi-k3",
  "messages": [
    {
      "role": "user",
      "content": "List three checks before saving a payment."
    }
  ],
  "max_completion_tokens": 2048,
  "reasoning_effort": "low",
  "stream": true
}'

SSE / stream=true

A successful stream has Content-Type: text/event-stream and X-Request-Id. Events end with a blank line; network chunk boundaries are not event boundaries. Collect data: lines to the end of an event before parsing JSON. Handle event: names, control blocks and errors; do not parse [DONE] as JSON. The shortened examples below use illustrative text and token counts.

event: response.output_text.delta
data: {"type":"response.output_text.delta","delta":"Ready."}

event: response.completed
data: {"type":"response.completed","response":{"id":"8488e6a4-d38a-4f95-a831-752aa12bf5d7","status":"completed","usage":{"input_tokens":12,"output_tokens":3}}}

Responses: accumulate response.output_text.delta.delta; completion events are response.completed or response.incomplete. The latter is partial output. Function-call arguments arrive in separate events and must also be assembled before JSON parsing.

data: {"choices":[{"index":0,"delta":{"content":"Ready."},"finish_reason":null}]}

data: {"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

data: {"choices":[],"usage":{"prompt_tokens":12,"completion_tokens":3,"total_tokens":15}}

data: [DONE]

Chat: accumulate choices[].delta by index; content and tool_calls may be partial. Do not stop reading at finish_reason: a usage chunk with empty choices may follow.

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Ready."}}

event: message_delta
data: {"type":"message_delta","delta":{"stop_reason":"end_turn"},"usage":{"output_tokens":3}}

event: message_stop
data: {"type":"message_stop"}

Messages: handle message_start, content_block_start/delta/stop, message_delta and message_stop. text_delta.text contains text; input_json_delta.partial_json contains partial tool arguments. Usage counters may arrive in multiple events.

Save X-Request-Id for support. A broken stream or client timeout does not cancel the request; partial text does not establish success. Do not automatically create a new paid operation after an interruption. Repeating a synchronous request with the same key returns 409 request_already_submitted rather than replaying the text. See the shared error reference.

wan-3

2–30 seconds; 480p / 720p / 1080p; up to 10 images, 5 videos, 5 audios, 20 total. Required prompt: at most 20000 characters. Frames and edit are supported. In edit omit duration, aspect_ratio and size: the first video determines dimensions. Seed 0–4294967295 and generate_audio: false are supported.

POST /v1/videos/generations
{
  "model": "wan-3",
  "prompt": "A slow camera move around a ceramic cup",
  "duration": 5,
  "resolution": "720p",
  "generate_audio": false,
  "seed": 42
}

POST /v1/videos/generations
{
  "model": "wan-3",
  "prompt": "Make the colors in @Video 1 warmer",
  "omni_reference_task_type": "edit",
  "resolution": "720p",
  "reference_videos": [
    {
      "url": "https://media.example.com/source.mp4"
    }
  ]
}

References are HTTPS URLs. No upfront hold is placed on partner funds. Confirmed usage.billed_seconds determines the final charge. Images and audio have no extra fee. One request produces one video. Save request_id from HTTP 202, poll GET /v1/videos/{request_id}; on done download /v1/videos/{request_id}/content. Do not replay a slow POST with a new key.

Input clips: 1–15 seconds, 15 total. Output and video references together: at most 30 seconds.

wan-3-prime

4–30 seconds; 480p / 720p / 1080p; up to 10 images, 5 videos, 5 audios, 20 total. Required prompt: at most 20000 characters. Ratios: 16:9, 1:1, 9:16. No frames, edit, seed, adaptive ratio or silent output. Audio references need an image or video.

POST /v1/videos/generations
{
  "model": "wan-3-prime",
  "prompt": "Animate the subject in @Image 1",
  "duration": 5,
  "resolution": "720p",
  "aspect_ratio": "16:9",
  "reference_images": [
    {
      "url": "https://media.example.com/reference.jpg"
    }
  ]
}

References are HTTPS URLs. No upfront hold is placed on partner funds. Confirmed usage.billed_seconds determines the final charge. Images and audio have no extra fee. One request produces one video. Save request_id from HTTP 202, poll GET /v1/videos/{request_id}; on done download /v1/videos/{request_id}/content. Do not replay a slow POST with a new key.

Input clips: 1–15 seconds, 15 total. Output and video references together: at most 30 seconds.