API reference¶

This task-oriented reference covers the routes implemented by Nexus. All examples use a placeholder key and a model returned for the authenticated key.

Common contract¶

Authentication: Authorization: Bearer <key>, x-api-key and api-key. Gemini routes also accept x-goog-api-key or ?key=.
Model placement: JSON body for OpenAI/Anthropic routes; {model} and an action suffix in a Gemini path; query model for Realtime.
Request ID: send x-request-id for correlation; Nexus returns it in the response header and local error envelope.
Limits: ordinary JSON routes accept up to 32 MiB; Gemini JSON up to 20 MiB; image/audio JSON up to 1 MiB; multipart image/audio up to 25 MiB.
Billing: successful measured usage or a compatible fixed-price response can be charged once. Missing or contradictory usage is left uncharged and is not charged later automatically.

export NEXUS_API_KEY='replace-with-your-key'
curl https://api.nexus-hub.ru/v1/models \
  -H "Authorization: Bearer $NEXUS_API_KEY" \
  -H "x-request-id: docs-models-001"

A successful discovery response has this shape:

{"object":"list","data":[{"id":"MODEL","object":"model","created":0,"owned_by":"..."}]}

Local failures use one safe envelope across route families:

{"error":{"message":"model is not enabled for this API key","type":"nexus_error","code":"model_not_allowed","request_id":"docs-001"}}

Discovery and account¶

Method and path	Model placement	Successful response
`GET /v1/models`	none	key-visible OpenAI-style list
`GET /v1/catalog`	none	public models and verified endpoint capabilities
`GET /v1/models/{model}`	path	one visible model or `model_not_found`
`GET /v1/balance`	none	balance object for the key's profile
`GET /v1beta/models`	none	Gemini aliases, methods and optional token limits
`GET /v1beta/openai/models`	none	OpenAI-style Gemini model list for bridge clients
`GET /v1/catalog?currency=RUB\\|USD\\|EUR\\|CNY`	none	public retail catalogue and price snapshot

Authenticated discovery routes are filtered by key, endpoint policy and active model access.

Public catalog¶

GET /v1/catalog needs no API key and defaults to RUB. It accepts USD, EUR, and CNY with the currency query parameter. An unsupported value returns 400 with invalid_catalog_currency.

The response keeps the order of the active public catalogue, families, public model IDs and display names, endpoint families, billing mode, and final retail price components. It is not a per-key availability response; use authenticated GET /v1/models for that. Each money value is a decimal string with an explicit unit such as USD_per_1M_tokens or EUR_per_request. snapshot.fx_status is current, stale, or unavailable; when it is unavailable, the affected components are null rather than values incorrectly labelled as the selected currency.

{"snapshot":{"fetched_at":"2026-08-14T10:00:00Z","effective_date":"2026-08-14","fx_status":"current"},"currency":"EUR","families":[{"id":"example","models":[{"id":"MODEL","display_name":"MODEL","endpoint_families":["responses"],"prices":[{"endpoint_families":["responses"],"billing_mode":"per_token","components":{"input":{"amount":"1.250000","unit":"EUR_per_1M_tokens"},"cached_input":null,"cache_write":null,"output":{"amount":"5.000000","unit":"EUR_per_1M_tokens"},"request":null}}]}]}]}

This route is intended for the documentation table. It supports browser cache revalidation with ETag and has a short public cache lifetime.

GET /v1/catalog is a no-auth capability discovery endpoint. It includes only public model metadata and verified optional parameters; missing endpoint entries remain unknown and do not become unsupported.

Text generation¶

All rows below accept JSON up to 32 MiB. Use a model returned for the key.

Method and path	Minimal JSON body	Successful response	Streaming/output control
`POST /v1/chat/completions`	`{"model":"MODEL","messages":[{"role":"user","content":"Reply OK"}],"max_completion_tokens":16}`	OpenAI Chat object	`stream: true`; `max_completion_tokens` or `max_tokens`
`POST /v1/completions`	`{"model":"MODEL","prompt":"Reply OK","max_tokens":16}`	OpenAI completion object	no stream; `max_tokens`
`POST /v1/responses`	`{"model":"MODEL","input":"Reply OK","max_output_tokens":16}`	Responses object	`stream: true`; `max_output_tokens`
`POST /v1/responses/compact`	`{"model":"MODEL","input":"Reply OK","max_output_tokens":16}`	compact Responses object	`stream: true`; `max_output_tokens`
`POST /v1/messages`	`{"model":"MODEL","max_tokens":16,"messages":[{"role":"user","content":"Reply OK"}]}`	Anthropic Messages object	`stream: true`; positive `max_tokens` required
`POST /v1/messages/count_tokens`	`{"model":"MODEL","messages":[{"role":"user","content":"Reply OK"}]}`	`{"input_tokens":...}`	no charge and no stream

Send anthropic-version: 2023-06-01 on both Messages routes. JSON streaming uses protocol-native SSE terminal events; a non-streaming request returns one JSON object. /v1/completions accepts only a string prompt: batch and token-ID prompts return unsupported_prompt_shape.

Minimal Responses request:

curl https://api.nexus-hub.ru/v1/responses \
  -H "Authorization: Bearer $NEXUS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"MODEL_FROM_V1_MODELS","input":"Reply exactly OK.","max_output_tokens":16}'

Minimal Messages request:

curl https://api.nexus-hub.ru/v1/messages \
  -H "x-api-key: $NEXUS_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{"model":"MODEL_FROM_V1_MODELS","max_tokens":16,"messages":[{"role":"user","content":"Reply exactly OK."}]}'

Embeddings, ranking and moderation¶

These JSON routes accept up to 32 MiB and do not stream.

Method and path	Minimal JSON body	Successful response
`POST /v1/embeddings`	`{"model":"MODEL","input":"Reply OK"}`	embedding list
`POST /v1/rerank`	`{"model":"MODEL","query":"OK","documents":["OK"],"top_n":1}`	ranked results
`POST /v1/moderations`	`{"model":"MODEL","input":"Reply OK"}`	moderation results

The key must have access to the endpoint family and model.

Images and audio¶

Method and path	Minimal request fields	Limit and successful response
`POST /v1/images/generations`	JSON: `model`, `prompt`	1 MiB; JSON image result
`POST /v1/images/edits`	multipart: `model`, `image`, `prompt`	25 MiB; JSON image result
`POST /v1/images/variations`	multipart: `model`, `image`	25 MiB; JSON image result
`POST /v1/audio/speech`	JSON: `model`, `input`, `voice`	1 MiB; binary audio
`POST /v1/audio/transcriptions`	multipart: `model`, `file`	25 MiB; JSON transcript
`POST /v1/audio/translations`	multipart: `model`, `file`	25 MiB; JSON translation

Media routes do not stream. A wrong content type returns unsupported_content_type; an oversized body returns payload_too_large.

Gemini-compatible actions¶

The path is POST /v1beta/models/{model}:{action}. All action bodies accept up to 20 MiB. The model is in the path; if a body also contains model, it must match.

Action	Minimal request shape	Response/stream and output control
`generateContent`	`contents`; optional `generationConfig`	JSON; `generationConfig.maxOutputTokens`
`streamGenerateContent`	same	SSE chunks, terminal `[DONE]`; same output field
`countTokens`	`contents`	`{"totalTokens":...}`; no charge and no stream
`embedContent`	`content`	one JSON embedding; no stream
`batchEmbedContents`	`requests[]`	JSON embeddings; no stream

curl "https://api.nexus-hub.ru/v1beta/models/MODEL:generateContent" \
  -H "x-goog-api-key: $NEXUS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents":[{"role":"user","parts":[{"text":"Reply exactly OK."}]}],"generationConfig":{"maxOutputTokens":16}}'

countTokens is a preflight operation and does not change the balance. embedContent accepts one content; batchEmbedContents accepts a requests array.

Body-size limits and output-token limits are separate. When Nexus has an explicit output limit for a model, a larger max_tokens, max_completion_tokens, max_output_tokens or Gemini generationConfig.maxOutputTokens is rejected with invalid_request_error. Gemini discovery exposes outputTokenLimit when that metadata is configured; an omitted value does not create a hidden default.

Realtime WebSocket¶

Connect to wss://api.nexus-hub.ru/v1/realtime?model=MODEL_FROM_V1_MODELS and send the key as Authorization: Bearer or x-api-key during the handshake. The query model is fixed for the session; a message attempting to switch it is rejected. Text, binary, ping, pong and close frames are proxied. Output limits inside protocol messages are forwarded unchanged.

Errors and retries¶

Common codes are invalid_key, model_not_found, model_not_allowed, endpoint_not_allowed, insufficient_balance, streaming_not_supported, unsupported_content_type, payload_too_large, upstream_error, upstream_timeout, pricing_unavailable and duplicate_request_id. Retry only errors marked transient by your client policy, and include the request ID when asking for help. Do not retry blindly when a response may already have been delivered. See Errors and support.