Model API

Base URL and authentication

Every model call goes to one host:

The origingame.ai site does not forward model requests, and browsers never relay model calls through the Portal. Call api.origingame.ai directly from your server or script.

Authenticate with the account API token (it starts with og-). Create, copy or rotate it at /me/access. An account has one; rotating it invalidates the old value at once. Signed-in devices and connected apps hold their own credentials and do not use this token.

The token acts as your account and spends your balance. Keep it in server-side environment variables; never put it in a web page or a game, or commit it to a repository.

Compatible protocols and SDKs

Point the official SDK's base URL at Origin and use your og- token as the API key.

OpenAI

Chat Completions (/v1/chat/completions) and Responses (POST /v1/responses) are supported. Responses has no WebSocket transport; use HTTP.

import OpenAI from 'openai'

const client = new OpenAI({
  baseURL: 'https://api.origingame.ai/v1',
  apiKey: process.env.OG_API_KEY,
})

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.origingame.ai/v1", api_key=os.environ["OG_API_KEY"])

Anthropic

Messages (/v1/messages) and the unbilled POST /v1/messages/count_tokens (returns {input_tokens, estimated}) are supported. The Anthropic SDK appends /v1 itself, so its base URL has no /v1.

import os
import anthropic

client = anthropic.Anthropic(base_url="https://api.origingame.ai", api_key=os.environ["OG_API_KEY"])
message = client.messages.create(
    model="MODEL_ID",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}],
)
print(message.content[0].text)

Replace MODEL_ID with an id from Listing models whose supported_endpoint_types contains anthropic.

Gemini

POST /v1beta/models/{model}:generateContent and :streamGenerateContent are supported; other Gemini actions return 404.

curl "https://api.origingame.ai/v1beta/models/MODEL_ID:generateContent" \
  -H "x-goog-api-key: $OG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents":[{"parts":[{"text":"Hello"}]}]}'

import os
from google import genai
from google.genai import types

client = genai.Client(
    api_key=os.environ["OG_API_KEY"],
    http_options=types.HttpOptions(base_url="https://api.origingame.ai"),
)
response = client.models.generate_content(model="MODEL_ID", contents="Hello")
print(response.text)

For MODEL_ID, pick a model whose supported_endpoint_types contains gemini.

Listing models

GET /v1/models returns the models your token can use. The catalog changes as operations configure it, and account and token model restrictions shape the result, so rely on the answer you get rather than hard-coding ids.

curl https://api.origingame.ai/v1/models \
  -H "Authorization: Bearer $OG_API_KEY"

The default answer is OpenAI-shaped ({"object":"list","data":[…]}). Each entry's supported_endpoint_types names the endpoints the model serves, for example openai, openai-response, anthropic, gemini, image-generation, audio-tts, openai-video. A request carrying both x-api-key and anthropic-version gets the Anthropic shape; one carrying x-goog-api-key gets the Gemini shape ({"models":[…]}).

Prices, capabilities and per-model integration guides with samples are in the model catalog; a single model is at /tools/models/<model id>.

For native GPT-Realtime-2.1 voice, endpoint metadata uses audio-realtime (GET /v1/realtime?model={model}), distinct from audio-stt-realtime (GET /v1/realtime?intent=transcription&model={model}) for standalone transcription.

Chat and streaming

The samples use gpt-5.6-sol. If your model list does not include it, use any id from the list whose supported_endpoint_types contains openai.

curl https://api.origingame.ai/v1/chat/completions \
  -H "Authorization: Bearer $OG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.6-sol","messages":[{"role":"user","content":"Introduce yourself in one sentence"}]}'

For streaming, add "stream": true; the response is SSE:

curl -N https://api.origingame.ai/v1/chat/completions \
  -H "Authorization: Bearer $OG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.6-sol","stream":true,"messages":[{"role":"user","content":"Write a short poem about pixels"}]}'

import OpenAI from 'openai'

const client = new OpenAI({
  baseURL: 'https://api.origingame.ai/v1',
  apiKey: process.env.OG_API_KEY,
})

const stream = await client.chat.completions.create({
  model: 'gpt-5.6-sol',
  messages: [{ role: 'user', content: 'Write a short poem about pixels' }],
  stream: true,
})
for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? '')
}

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.origingame.ai/v1", api_key=os.environ["OG_API_KEY"])

stream = client.chat.completions.create(
    model="gpt-5.6-sol",
    messages=[{"role": "user", "content": "Write a short poem about pixels"}],
    stream=True,
)
for chunk in stream:
    if chunk.choices and chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

Endpoint reference

Every path is under https://api.origingame.ai; authentication is described in Base URL and authentication. Requests and responses match each vendor's native API; per-model request samples are on the model pages of the model catalog.

Text

Images

Speech and audio

The live transcription WebSocket needs an Authorization header on the handshake, which browsers cannot set, so connect from a server.

Native GPT-Realtime-2.1

This describes the current workspace implementation, not a claim of production deployment or full verification. The GA native protocol passes through events including concurrent/out-of-band responses, VAD, tools, images, audio codecs and interruptions. Sessions are capped at 60 minutes. The 32 MiB serialized WebSocket message ceiling is a Gateway limit, not a universal upstream bound. Legacy beta session endpoints are not implemented. The strict intent=transcription GPT-Live path above remains separate from native voice and its input ASR.

Client capabilities last 10–7200 seconds, are reusable, and revalidate the parent credential on use; no provider key is returned. For SIP, operators configure the root runtime secret realtime.sip_bindings array (route, project_id, channel_id, user_id, webhook_secret). Signed inbox bindings establish the tenant, never phone headers. Users receive the original OpenAI event through a separate project webhook registration. Real SIP acceptance needs a provisioned project webhook, trunk and test phone; these are not yet available or verified.

Optional X-Origin-Realtime-Budget specifies a pre-funded full-session cap in USD (conservative default: $6.144 for gpt-realtime-2.1). Prices are frozen for the session. Authoritative terminal response and ASR usage are accounted for separately, with customer settlement capped at the funded amount. The platform bears excess, unpriced or unobserved usage. Termination is requested on loss of observation or budget protection; it is not guaranteed instantaneous. Client disconnect briefly drains cancellation usage; all closes conservatively mark observation incomplete rather than inventing usage.

Video

3D

Decisions and classification

Media files

Async tasks

Video, speech tasks and 3D are asynchronous: the create request returns a task object at once (usually 202, 200 on some video channels); you then poll the status endpoint and download the result when it is done.

1. Create the task and keep the returned id. Where Idempotency-Key is required, generate a unique value for each new task.
2. Query .../{id} every few seconds until status is completed or failed.
3. After completed, download the result from result_url or .../{id}/content.

curl https://api.origingame.ai/v1/tripo/tasks \
  -H "Authorization: Bearer $OG_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $(uuidgen)" \
  -d '{"model":"tripo-v3.1","operation":"text","parameters":{"prompt":"A low-poly treasure chest","texture":true,"pbr":true}}'

curl https://api.origingame.ai/v1/tripo/tasks/TASK_ID \
  -H "Authorization: Bearer $OG_API_KEY"

curl https://api.origingame.ai/v1/tripo/tasks/TASK_ID/content \
  -H "Authorization: Bearer $OG_API_KEY" -o model.bin

Common states: queued, in_progress, completed, failed, submission_unknown.

• Resending the same request with the same Idempotency-Key returns the original task instead of creating another; the same key with different parameters returns 409.
• submission_unknown means it is not known whether the provider accepted the job. Do not resubmit under a new key, or you may pay twice; the task's hold is kept for the platform to reconcile.
• Generated files are kept for the platform's retention period; copy them to your own storage promptly.

Errors

Errors keep the shape of the protocol you called. OpenAI-compatible and Gemini endpoints:

{"error":{"message":"...","type":"new_api_error","param":"","code":"..."}}

/v1/messages answers in the Anthropic shape: {"type":"error","error":{"type":"...","message":"..."}}. An unknown path is 404 with error.code not_found.

Environment variables

The platform's scripts and skills use these two variables; reuse the same names in your own code:

export OG_API_KEY=og-...
export OG_AI_GATEWAY=https://api.origingame.ai
curl "$OG_AI_GATEWAY/v1/models" -H "Authorization: Bearer $OG_API_KEY"