Calling OpenRouter from Python and TypeScript SDKs

Q: How do I send provider.order with TypeScript?

Include a provider object on the create call (may need a type assertion) matching OpenRouter's JSON schema.

Q: Is Python 3.9 enough?

Use the Python version required by your pinned openai package (3.11+ recommended).

Q: What about structured outputs?

Use OpenRouter/model-supported response_format or tool schemas, then validate with Pydantic/Zod in your code regardless of provider.

Most agent codebases already depend on the OpenAI client libraries.

You can keep those libraries and call hundreds of models by changing base URL, key, and model slug - then pass OpenRouter-only options through extension fields when needed.

Summary

Use the official OpenAI Python and TypeScript SDKs pointed at https://openrouter.ai/api/v1.

Send standard chat.completions (or streaming) requests with OpenRouter model slugs.

Use extra_body (Python) or extra request fields (TypeScript) for OpenRouter extensions such as models fallbacks and provider preferences.

Recipe

One client per process; one function that accepts messages + logical model config.

import os
from openai import OpenAI
 
client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key=os.environ["OPENROUTER_API_KEY"],
)
 
def chat(model: str, messages: list[dict], **openrouter_extra):
    return client.chat.completions.create(
        model=model,
        messages=messages,
        extra_body=openrouter_extra or None,
    )

When to reach for this:

Python services and notebooks
Node/TypeScript agents and Next.js route handlers (server-side only)
Shared helpers used by LangGraph nodes, tools, or eval harnesses
Teams standardizing on the OpenAI SDK instead of raw HTTP

Working Example

Python - non-streaming completion

import os
from openai import OpenAI
 
client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key=os.environ["OPENROUTER_API_KEY"],
)
 
resp = client.chat.completions.create(
    model="anthropic/claude-sonnet-4.5",  # verify at build
    messages=[
        {"role": "system", "content": "Answer in one short paragraph."},
        {"role": "user", "content": "What is model routing?"},
    ],
    temperature=0.2,
    extra_headers={
        "HTTP-Referer": "https://example.com",
        "X-OpenRouter-Title": "sdk-recipe",
    },
    extra_body={
        "models": ["openai/gpt-4o-mini"],  # fallbacks; verify at build
        "provider": {"sort": "throughput"},
    },
)
 
print(resp.model)
print(resp.choices[0].message.content)
print(resp.usage)

TypeScript - non-streaming completion

import OpenAI from "openai";
 
const client = new OpenAI({
  baseURL: "https://openrouter.ai/api/v1",
  apiKey: process.env.OPENROUTER_API_KEY,
  defaultHeaders: {
    "HTTP-Referer": "https://example.com",
    "X-OpenRouter-Title": "sdk-recipe",
  },
});
 
const resp = await client.chat.completions.create({
  model: "anthropic/claude-sonnet-4.5", // verify at build
  messages: [
    { role: "system", content: "Answer in one short paragraph." },
    { role: "user", content: "What is model routing?" },
  ],
  temperature: 0.2,
  // OpenRouter extensions (types may not include them):
  // @ts-expect-error OpenRouter-only fields
  models: ["openai/gpt-4o-mini"],
  // @ts-expect-error OpenRouter-only fields
  provider: { sort: "throughput" },
});
 
console.log(resp.model);
console.log(resp.choices[0].message.content);
console.log(resp.usage);

Python - streaming deltas

stream = client.chat.completions.create(
    model="openai/gpt-4o-mini",  # verify at build
    messages=[{"role": "user", "content": "List three agent tools."}],
    stream=True,
)
for chunk in stream:
    text = chunk.choices[0].delta.content or ""
    if text:
        print(text, end="", flush=True)

TypeScript - streaming deltas

const stream = await client.chat.completions.create({
  model: "openai/gpt-4o-mini", // verify at build
  messages: [{ role: "user", content: "List three agent tools." }],
  stream: true,
});
 
for await (const chunk of stream) {
  const text = chunk.choices[0]?.delta?.content ?? "";
  if (text) process.stdout.write(text);
}

What this demonstrates:

Identical product behavior from Python and TypeScript
Standard OpenAI parameters still work (messages, temperature, stream)
OpenRouter extensions ride alongside without a second HTTP client
Resolved model and usage are available after the call

Deep Dive

Install

# Python
pip install openai
 
# TypeScript / Node
npm install openai

Pin versions in your lockfile; verify current major at build.

Parameter map

Goal	OpenAI SDK field	OpenRouter note
Host	`base_url` / `baseURL`	`https://openrouter.ai/api/v1`
Auth	`api_key` / `apiKey`	OpenRouter key
Model	`model`	`provider/model` slug
Fallbacks	`extra_body.models` (Py) / `models` (TS)	Ordered backups
Provider prefs	`extra_body.provider`	order, sort, ignore, zdr, ...
Streaming	`stream=True` / `stream: true`	SSE under the hood
Attribution	`extra_headers` / `defaultHeaders`	Optional

Tool calls

Function calling works through the same Chat Completions shapes when the selected model/provider supports tools.

OpenRouter routes tool requests toward capable endpoints when possible (verify tool-routing behavior at build).

Always re-test tool schemas after changing model policy.

Official OpenRouter SDKs (optional)

# pip install openrouter  (verify package name at build)
from openrouter import OpenRouter
import os
 
with OpenRouter(api_key=os.environ["OPENROUTER_API_KEY"]) as client:
    response = client.chat.send(
        model="openai/gpt-4o-mini",  # verify at build
        messages=[{"role": "user", "content": "Hello"}],
    )
    print(response.choices[0].message.content)

// npm install @openrouter/sdk  (verify package at build)
import { OpenRouter } from "@openrouter/sdk";
 
const openrouter = new OpenRouter({
  apiKey: process.env.OPENROUTER_API_KEY!,
});
 
const completion = await openrouter.chat.send({
  model: "openai/gpt-4o-mini", // verify at build
  messages: [{ role: "user", content: "Hello" }],
});
console.log(completion.choices[0].message.content);

Prefer these when you want first-class types for OpenRouter-only features.

Prefer the OpenAI SDK when the rest of the monorepo already standardizes on it.

Error handling sketch

from openai import APIStatusError
 
try:
    resp = client.chat.completions.create(
        model="openai/gpt-4o-mini",
        messages=[{"role": "user", "content": "hi"}],
    )
except APIStatusError as e:
    # 401 auth, 402 credits, 429 rate limit, 503 no provider, ...
    print(e.status_code, e.response.text if e.response is not None else e.message)
    raise

Honor Retry-After on 429/503 when present.

Gotchas

Bare OpenAI model ids without provider prefix. OpenRouter expects catalog slugs such as openai/gpt-4o-mini. Fix: use full slugs from openrouter.ai/models.
Passing OpenRouter fields without extra_body in Python. They may be dropped or rejected. Fix: put models / provider in extra_body.
TypeScript compile errors on extension fields. Types lag product. Fix: narrow cast or @ts-expect-error with a comment.
Browser bundles with secret keys. Keys leak. Fix: server-only modules.
Assuming tool-call message shapes are identical across every model. Providers differ. Fix: golden tool-call tests per policy tier.
Ignoring resp.model after fallbacks. Cost and quality retros break. Fix: log requested vs resolved model every time.
Mixing Responses API beta and Chat Completions casually. Different skins. Fix: stick to one API shape per code path until you need the other.

Alternatives

Alternative	Use When	Don't Use When
OpenAI SDK (this page)	Shared OpenAI ecosystem	You need every OpenRouter admin API typed
Official OpenRouter SDK	Greenfield OpenRouter apps	Team forbids new deps
Vercel AI SDK + OpenRouter provider	Web UI streaming stacks	Pure Python workers
LangChain / LlamaIndex chat models	Framework agents	You only need a 20-line client
Raw HTTP	Debugging	Large production surface without retries

FAQs

Which SDK version should I pin?

Pin the latest stable OpenAI SDK your framework supports and re-verify OpenRouter compatibility at build.

Can I use AsyncOpenAI?

Yes.

from openai import AsyncOpenAI
client = AsyncOpenAI(base_url="https://openrouter.ai/api/v1", api_key=...)

How do I send provider.order with TypeScript?

Include a provider object on the create call (may need a type assertion) matching OpenRouter's JSON schema.

Does `max_tokens` still work?

Usually yes as a Chat Completions parameter.

Some models use related fields; check OpenRouter parameter docs for the model family at build.

How do I read the generation id?

OpenRouter may expose generation identifiers via response fields or headers such as X-Generation-Id (verify at build).

Log them for support tickets.

Can one client instance call many models?

Yes.

Model is per-request; base URL and key stay fixed.

Is Python 3.9 enough?

Use the Python version required by your pinned openai package (3.11+ recommended).

How do embeddings work?

OpenRouter exposes embeddings endpoints separately.

Configure the same base URL/key, then call the embeddings API for supported models (verify catalog at build).

Should eval harnesses use the same helper as production?

Yes.

Shared clients keep slug and header bugs out of "tests green, prod red" traps.

What about structured outputs?

Use OpenRouter/model-supported response_format or tool schemas, then validate with Pydantic/Zod in your code regardless of provider.

Authenticating and Configuring the OpenRouter Base URL - client construction detail
Streaming Responses Through OpenRouter - stream edge cases
Model Slugs and Naming Conventions on OpenRouter - slug reference
Switching Models by Config, Not Code: A Routing Pattern - keep slugs out of call sites
OpenRouter Setup & Routing Basics - minimal first call

Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana @solana/web3.js + @solana/spl-token.

You can keep those libraries and call hundreds of models by changing base URL, key, and model slug - then pass OpenRouter-only options through extension fields when needed.

Summary

Use the official OpenAI Python and TypeScript SDKs pointed at https://openrouter.ai/api/v1.

Send standard chat.completions (or streaming) requests with OpenRouter model slugs.

Use extra_body (Python) or extra request fields (TypeScript) for OpenRouter extensions such as models fallbacks and provider preferences.

Recipe

One client per process; one function that accepts messages + logical model config.

import os
from openai import OpenAI
 
client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key=os.environ["OPENROUTER_API_KEY"],
)
 
def chat(model: str, messages: list[dict], **openrouter_extra):
    return client.chat.completions.create(
        model=model,
        messages=messages,
        extra_body=openrouter_extra or None,
    )

When to reach for this:

Python services and notebooks
Node/TypeScript agents and Next.js route handlers (server-side only)
Shared helpers used by LangGraph nodes, tools, or eval harnesses
Teams standardizing on the OpenAI SDK instead of raw HTTP

Working Example

Python - non-streaming completion

import os
from openai import OpenAI
 
client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key=os.environ["OPENROUTER_API_KEY"],
)
 
resp = client.chat.completions.create(
    model="anthropic/claude-sonnet-4.5",  # verify at build
    messages=[
        {"role": "system", "content": "Answer in one short paragraph."},
        {"role": "user", "content": "What is model routing?"},
    ],
    temperature=0.2,
    extra_headers={
        "HTTP-Referer": "https://example.com",
        "X-OpenRouter-Title": "sdk-recipe",
    },
    extra_body={
        "models": ["openai/gpt-4o-mini"],  # fallbacks; verify at build
        "provider": {"sort": "throughput"},
    },
)
 
print(resp.model)
print(resp.choices[0].message.content)
print(resp.usage)

TypeScript - non-streaming completion

import OpenAI from "openai";
 
const client = new OpenAI({
  baseURL: "https://openrouter.ai/api/v1",
  apiKey: process.env.OPENROUTER_API_KEY,
  defaultHeaders: {
    "HTTP-Referer": "https://example.com",
    "X-OpenRouter-Title": "sdk-recipe",
  },
});
 
const resp = await client.chat.completions.create({
  model: "anthropic/claude-sonnet-4.5", // verify at build
  messages: [
    { role: "system", content: "Answer in one short paragraph." },
    { role: "user", content: "What is model routing?" },
  ],
  temperature: 0.2,
  // OpenRouter extensions (types may not include them):
  // @ts-expect-error OpenRouter-only fields
  models: ["openai/gpt-4o-mini"],
  // @ts-expect-error OpenRouter-only fields
  provider: { sort: "throughput" },
});
 
console.log(resp.model);
console.log(resp.choices[0].message.content);
console.log(resp.usage);

Python - streaming deltas

stream = client.chat.completions.create(
    model="openai/gpt-4o-mini",  # verify at build
    messages=[{"role": "user", "content": "List three agent tools."}],
    stream=True,
)
for chunk in stream:
    text = chunk.choices[0].delta.content or ""
    if text:
        print(text, end="", flush=True)

TypeScript - streaming deltas

const stream = await client.chat.completions.create({
  model: "openai/gpt-4o-mini", // verify at build
  messages: [{ role: "user", content: "List three agent tools." }],
  stream: true,
});
 
for await (const chunk of stream) {
  const text = chunk.choices[0]?.delta?.content ?? "";
  if (text) process.stdout.write(text);
}

What this demonstrates:

Identical product behavior from Python and TypeScript
Standard OpenAI parameters still work (messages, temperature, stream)
OpenRouter extensions ride alongside without a second HTTP client
Resolved model and usage are available after the call

Deep Dive

Install

# Python
pip install openai
 
# TypeScript / Node
npm install openai

Pin versions in your lockfile; verify current major at build.

Parameter map

Goal	OpenAI SDK field	OpenRouter note
Host	`base_url` / `baseURL`	`https://openrouter.ai/api/v1`
Auth	`api_key` / `apiKey`	OpenRouter key
Model	`model`	`provider/model` slug
Fallbacks	`extra_body.models` (Py) / `models` (TS)	Ordered backups
Provider prefs	`extra_body.provider`	order, sort, ignore, zdr, ...
Streaming	`stream=True` / `stream: true`	SSE under the hood
Attribution	`extra_headers` / `defaultHeaders`	Optional

Tool calls

Function calling works through the same Chat Completions shapes when the selected model/provider supports tools.

OpenRouter routes tool requests toward capable endpoints when possible (verify tool-routing behavior at build).

Always re-test tool schemas after changing model policy.

Official OpenRouter SDKs (optional)

# pip install openrouter  (verify package name at build)
from openrouter import OpenRouter
import os
 
with OpenRouter(api_key=os.environ["OPENROUTER_API_KEY"]) as client:
    response = client.chat.send(
        model="openai/gpt-4o-mini",  # verify at build
        messages=[{"role": "user", "content": "Hello"}],
    )
    print(response.choices[0].message.content)

// npm install @openrouter/sdk  (verify package at build)
import { OpenRouter } from "@openrouter/sdk";
 
const openrouter = new OpenRouter({
  apiKey: process.env.OPENROUTER_API_KEY!,
});
 
const completion = await openrouter.chat.send({
  model: "openai/gpt-4o-mini", // verify at build
  messages: [{ role: "user", content: "Hello" }],
});
console.log(completion.choices[0].message.content);

Prefer these when you want first-class types for OpenRouter-only features.

Prefer the OpenAI SDK when the rest of the monorepo already standardizes on it.

Error handling sketch

from openai import APIStatusError
 
try:
    resp = client.chat.completions.create(
        model="openai/gpt-4o-mini",
        messages=[{"role": "user", "content": "hi"}],
    )
except APIStatusError as e:
    # 401 auth, 402 credits, 429 rate limit, 503 no provider, ...
    print(e.status_code, e.response.text if e.response is not None else e.message)
    raise

Honor Retry-After on 429/503 when present.

Gotchas

Bare OpenAI model ids without provider prefix. OpenRouter expects catalog slugs such as openai/gpt-4o-mini. Fix: use full slugs from openrouter.ai/models.
Passing OpenRouter fields without extra_body in Python. They may be dropped or rejected. Fix: put models / provider in extra_body.
TypeScript compile errors on extension fields. Types lag product. Fix: narrow cast or @ts-expect-error with a comment.
Browser bundles with secret keys. Keys leak. Fix: server-only modules.
Assuming tool-call message shapes are identical across every model. Providers differ. Fix: golden tool-call tests per policy tier.
Ignoring resp.model after fallbacks. Cost and quality retros break. Fix: log requested vs resolved model every time.
Mixing Responses API beta and Chat Completions casually. Different skins. Fix: stick to one API shape per code path until you need the other.

Alternatives

Alternative	Use When	Don't Use When
OpenAI SDK (this page)	Shared OpenAI ecosystem	You need every OpenRouter admin API typed
Official OpenRouter SDK	Greenfield OpenRouter apps	Team forbids new deps
Vercel AI SDK + OpenRouter provider	Web UI streaming stacks	Pure Python workers
LangChain / LlamaIndex chat models	Framework agents	You only need a 20-line client
Raw HTTP	Debugging	Large production surface without retries

FAQs

Which SDK version should I pin?

Pin the latest stable OpenAI SDK your framework supports and re-verify OpenRouter compatibility at build.

Can I use AsyncOpenAI?

Yes.

from openai import AsyncOpenAI
client = AsyncOpenAI(base_url="https://openrouter.ai/api/v1", api_key=...)

How do I send provider.order with TypeScript?

Include a provider object on the create call (may need a type assertion) matching OpenRouter's JSON schema.

Does `max_tokens` still work?

Usually yes as a Chat Completions parameter.

Some models use related fields; check OpenRouter parameter docs for the model family at build.

How do I read the generation id?

OpenRouter may expose generation identifiers via response fields or headers such as X-Generation-Id (verify at build).

Log them for support tickets.

Can one client instance call many models?

Yes.

Model is per-request; base URL and key stay fixed.

Is Python 3.9 enough?

Use the Python version required by your pinned openai package (3.11+ recommended).

How do embeddings work?

OpenRouter exposes embeddings endpoints separately.

Configure the same base URL/key, then call the embeddings API for supported models (verify catalog at build).

Should eval harnesses use the same helper as production?

Yes.

Shared clients keep slug and header bugs out of "tests green, prod red" traps.

What about structured outputs?

Use OpenRouter/model-supported response_format or tool schemas, then validate with Pydantic/Zod in your code regardless of provider.

Authenticating and Configuring the OpenRouter Base URL - client construction detail
Streaming Responses Through OpenRouter - stream edge cases
Model Slugs and Naming Conventions on OpenRouter - slug reference
Switching Models by Config, Not Code: A Routing Pattern - keep slugs out of call sites
OpenRouter Setup & Routing Basics - minimal first call

Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana @solana/web3.js + @solana/spl-token.

Calling OpenRouter from Python and TypeScript SDKs

Search Documentation

Calling OpenRouter from Python and TypeScript SDKs