Cost & Fallback Basics

8 examples to get you started with OpenRouter cost control and provider fallback - 5 basic and 3 intermediate.

You will call the chat API, set provider order, enable or disable fallbacks, sort by price, read usage cost, and sketch a per-key spend cap.

Prerequisites

Python 3.11+ recommended
OpenRouter account with credits and an API key
Optional: a Management API key if you create keys with credit limits via API

python -m venv .venv && source .venv/bin/activate
pip install openai requests
export OPENROUTER_API_KEY="sk-or-v1-..."
# Optional management key for /api/v1/keys (not for chat):
# export OPENROUTER_MANAGEMENT_KEY="sk-or-v1-..."

Basic Examples

1. Call OpenRouter Once and Read Cost

Start with a single completion so you see the usage object OpenRouter always returns.

import os
from openai import OpenAI
 
client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key=os.environ["OPENROUTER_API_KEY"],
    default_headers={
        "HTTP-Referer": "https://localhost",  # optional app attribution
        "X-OpenRouter-Title": "cost-fallback-basics",
    },
)
 
resp = client.chat.completions.create(
    model="openai/gpt-4o-mini",  # verify slug at build
    messages=[{"role": "user", "content": "Reply with one word: pong"}],
)
print(resp.choices[0].message.content)
print(resp.usage)  # includes cost when provided by OpenRouter

One request in, text plus usage out.
No provider prefs yet; OpenRouter load-balances for price and uptime.
Treat usage.cost as the amount charged for that generation (verify field names at build).

Related: Monitoring OpenRouter Spend and Usage in Real Time

2. Prefer a Provider Order With Fallbacks

Pin preferred hosts, then allow others if those fail.

resp = client.chat.completions.create(
    model="meta-llama/llama-3.3-70b-instruct",
    messages=[{"role": "user", "content": "Name two colors."}],
    extra_body={
        "provider": {
            "order": ["together", "fireworks", "deepinfra"],
            "allow_fallbacks": True,
        }
    },
)
print(resp.choices[0].message.content)

order is tried first, in list sequence.
With allow_fallbacks: True (default), non-listed providers may still serve as backups.
Copy exact provider slugs from the model page when you need variants like deepinfra/turbo.

Related: Setting Provider Order and Routing Preferences

3. Disable Fallbacks for a Strict Path

Fail closed when only your ordered providers may answer.

resp = client.chat.completions.create(
    model="meta-llama/llama-3.3-70b-instruct",
    messages=[{"role": "user", "content": "Say hi"}],
    extra_body={
        "provider": {
            "order": ["together", "fireworks"],
            "allow_fallbacks": False,
        }
    },
)

If both listed providers fail, the request errors instead of wandering.
Use this for compliance drills, not for maximum uptime.
Pair with monitoring so you notice the stricter error rate.

4. Sort Providers by Price (`:floor` Shortcut)

Force lowest-price routing for a slug without writing a full provider object.

resp = client.chat.completions.create(
    model="meta-llama/llama-3.3-70b-instruct:floor",
    messages=[{"role": "user", "content": "Summarize fallback in 8 words."}],
)
print(resp.choices[0].message.content)
# Equivalent: extra_body={"provider": {"sort": "price"}}

:floor sets provider.sort to "price".
Load balancing is disabled when you set sort or order.
Cheap is not always best for tool-calling quality; measure your agent tasks.

Related: Cost-Based Routing: Sending Cheap Tasks to Cheap Models

5. Check Key Credits Before a Batch

Poll GET /api/v1/key so a long agent job does not die mid-run on 402.

import requests
 
r = requests.get(
    "https://openrouter.ai/api/v1/key",
    headers={"Authorization": f"Bearer {os.environ['OPENROUTER_API_KEY']}"},
)
r.raise_for_status()
info = r.json()["data"]
print("limit:", info.get("limit"))
print("limit_remaining:", info.get("limit_remaining"))
print("usage:", info.get("usage"), "daily:", info.get("usage_daily"))

limit / limit_remaining describe an optional per-key credit cap.
Account balance and key cap both can produce 402 Payment Required.
Log these fields in CI or a cron before heavy eval suites.

Related: Setting Spend Limits and Usage Caps per API Key

Intermediate Examples

6. Combine Provider Order With Model Fallbacks

Survive both host failure and full model unavailability.

resp = client.chat.completions.create(
    model="anthropic/claude-sonnet-4.5",  # verify at build
    messages=[{"role": "user", "content": "Outline a 3-step plan."}],
    extra_body={
        "models": [
            "anthropic/claude-sonnet-4.5",
            "openai/gpt-4o-mini",
            "google/gemini-2.5-flash",  # verify slugs at build
        ],
        "provider": {
            "allow_fallbacks": True,
            "data_collection": "deny",  # optional privacy filter
        },
    },
)
print("served model:", resp.model)
print("usage:", resp.usage)

Provider hops try to keep the primary model.
The models list walks alternate models when the primary cannot serve.
You are billed for the model that actually ran.

7. Cap a New Key With the Management API

Create a customer or staging key with a hard dollar limit (management key required).

import os
import requests
 
mgmt = os.environ["OPENROUTER_MANAGEMENT_KEY"]
resp = requests.post(
    "https://openrouter.ai/api/v1/keys/",
    headers={
        "Authorization": f"Bearer {mgmt}",
        "Content-Type": "application/json",
    },
    json={
        "name": "agent-staging",
        "limit": 25.0,  # credit limit in USD-equivalent credits
        "limit_reset": "monthly",  # or daily / weekly / null; verify enums at build
    },
)
resp.raise_for_status()
print(resp.json())  # includes key secret once on create; store securely

Management keys cannot call chat completions.
Use separate keys per agent environment (dev / staging / prod).
Reset policies contain runaway loops without wiping the whole account.

8. Tiny Cost-Aware Agent Step

Route classification to a cheap model and hard reasoning to a stronger one.

def complete(model: str, prompt: str, **provider):
    return client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
        extra_body={"provider": provider} if provider else None,
    )
 
def agent_step(user_text: str) -> str:
    route = complete(
        "openai/gpt-4o-mini",
        f"Classify as CHEAP or HARD (one word): {user_text}",
        sort="price",
    ).choices[0].message.content or ""
    model = (
        "openai/gpt-4o-mini"
        if "CHEAP" in route.upper()
        else "anthropic/claude-sonnet-4.5"
    )
    out = complete(model, user_text, allow_fallbacks=True)
    print("route:", route, "model:", out.model, "usage:", out.usage)
    return out.choices[0].message.content or ""
 
print(agent_step("What is 2+2?"))

Application routing multiplies savings beyond provider sort alone.
Keep the classifier prompt short; it still costs tokens.
Always retain fallbacks on the expensive path so HARD tasks still complete.

Related: How OpenRouter's Provider Fallback Actually Protects Uptime

What to Practice Next

Tighten only / ignore lists once you know which providers pass evals.
Add latency sort for interactive agents.
Wire spend alerts from usage.cost and the Activity or Analytics APIs.

Related: Comparing OpenRouter's Per-Model Pricing Against Direct Provider Rates

Related: OpenRouter Cost & Fallback Best Practices

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.

Cost & Fallback Basics

8 examples to get you started with OpenRouter cost control and provider fallback - 5 basic and 3 intermediate.

You will call the chat API, set provider order, enable or disable fallbacks, sort by price, read usage cost, and sketch a per-key spend cap.

Prerequisites

Python 3.11+ recommended
OpenRouter account with credits and an API key
Optional: a Management API key if you create keys with credit limits via API

python -m venv .venv && source .venv/bin/activate
pip install openai requests
export OPENROUTER_API_KEY="sk-or-v1-..."
# Optional management key for /api/v1/keys (not for chat):
# export OPENROUTER_MANAGEMENT_KEY="sk-or-v1-..."

Basic Examples

1. Call OpenRouter Once and Read Cost

Start with a single completion so you see the usage object OpenRouter always returns.

import os
from openai import OpenAI
 
client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key=os.environ["OPENROUTER_API_KEY"],
    default_headers={
        "HTTP-Referer": "https://localhost",  # optional app attribution
        "X-OpenRouter-Title": "cost-fallback-basics",
    },
)
 
resp = client.chat.completions.create(
    model="openai/gpt-4o-mini",  # verify slug at build
    messages=[{"role": "user", "content": "Reply with one word: pong"}],
)
print(resp.choices[0].message.content)
print(resp.usage)  # includes cost when provided by OpenRouter

One request in, text plus usage out.
No provider prefs yet; OpenRouter load-balances for price and uptime.
Treat usage.cost as the amount charged for that generation (verify field names at build).

Related: Monitoring OpenRouter Spend and Usage in Real Time

2. Prefer a Provider Order With Fallbacks

Pin preferred hosts, then allow others if those fail.

resp = client.chat.completions.create(
    model="meta-llama/llama-3.3-70b-instruct",
    messages=[{"role": "user", "content": "Name two colors."}],
    extra_body={
        "provider": {
            "order": ["together", "fireworks", "deepinfra"],
            "allow_fallbacks": True,
        }
    },
)
print(resp.choices[0].message.content)

order is tried first, in list sequence.
With allow_fallbacks: True (default), non-listed providers may still serve as backups.
Copy exact provider slugs from the model page when you need variants like deepinfra/turbo.

Related: Setting Provider Order and Routing Preferences

3. Disable Fallbacks for a Strict Path

Fail closed when only your ordered providers may answer.

resp = client.chat.completions.create(
    model="meta-llama/llama-3.3-70b-instruct",
    messages=[{"role": "user", "content": "Say hi"}],
    extra_body={
        "provider": {
            "order": ["together", "fireworks"],
            "allow_fallbacks": False,
        }
    },
)

If both listed providers fail, the request errors instead of wandering.
Use this for compliance drills, not for maximum uptime.
Pair with monitoring so you notice the stricter error rate.

4. Sort Providers by Price (`:floor` Shortcut)

Force lowest-price routing for a slug without writing a full provider object.

resp = client.chat.completions.create(
    model="meta-llama/llama-3.3-70b-instruct:floor",
    messages=[{"role": "user", "content": "Summarize fallback in 8 words."}],
)
print(resp.choices[0].message.content)
# Equivalent: extra_body={"provider": {"sort": "price"}}

:floor sets provider.sort to "price".
Load balancing is disabled when you set sort or order.
Cheap is not always best for tool-calling quality; measure your agent tasks.

Related: Cost-Based Routing: Sending Cheap Tasks to Cheap Models

5. Check Key Credits Before a Batch

Poll GET /api/v1/key so a long agent job does not die mid-run on 402.

import requests
 
r = requests.get(
    "https://openrouter.ai/api/v1/key",
    headers={"Authorization": f"Bearer {os.environ['OPENROUTER_API_KEY']}"},
)
r.raise_for_status()
info = r.json()["data"]
print("limit:", info.get("limit"))
print("limit_remaining:", info.get("limit_remaining"))
print("usage:", info.get("usage"), "daily:", info.get("usage_daily"))

limit / limit_remaining describe an optional per-key credit cap.
Account balance and key cap both can produce 402 Payment Required.
Log these fields in CI or a cron before heavy eval suites.

Related: Setting Spend Limits and Usage Caps per API Key

Intermediate Examples

6. Combine Provider Order With Model Fallbacks

Survive both host failure and full model unavailability.

resp = client.chat.completions.create(
    model="anthropic/claude-sonnet-4.5",  # verify at build
    messages=[{"role": "user", "content": "Outline a 3-step plan."}],
    extra_body={
        "models": [
            "anthropic/claude-sonnet-4.5",
            "openai/gpt-4o-mini",
            "google/gemini-2.5-flash",  # verify slugs at build
        ],
        "provider": {
            "allow_fallbacks": True,
            "data_collection": "deny",  # optional privacy filter
        },
    },
)
print("served model:", resp.model)
print("usage:", resp.usage)

Provider hops try to keep the primary model.
The models list walks alternate models when the primary cannot serve.
You are billed for the model that actually ran.

7. Cap a New Key With the Management API

Create a customer or staging key with a hard dollar limit (management key required).

import os
import requests
 
mgmt = os.environ["OPENROUTER_MANAGEMENT_KEY"]
resp = requests.post(
    "https://openrouter.ai/api/v1/keys/",
    headers={
        "Authorization": f"Bearer {mgmt}",
        "Content-Type": "application/json",
    },
    json={
        "name": "agent-staging",
        "limit": 25.0,  # credit limit in USD-equivalent credits
        "limit_reset": "monthly",  # or daily / weekly / null; verify enums at build
    },
)
resp.raise_for_status()
print(resp.json())  # includes key secret once on create; store securely

Management keys cannot call chat completions.
Use separate keys per agent environment (dev / staging / prod).
Reset policies contain runaway loops without wiping the whole account.

8. Tiny Cost-Aware Agent Step

Route classification to a cheap model and hard reasoning to a stronger one.

def complete(model: str, prompt: str, **provider):
    return client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
        extra_body={"provider": provider} if provider else None,
    )
 
def agent_step(user_text: str) -> str:
    route = complete(
        "openai/gpt-4o-mini",
        f"Classify as CHEAP or HARD (one word): {user_text}",
        sort="price",
    ).choices[0].message.content or ""
    model = (
        "openai/gpt-4o-mini"
        if "CHEAP" in route.upper()
        else "anthropic/claude-sonnet-4.5"
    )
    out = complete(model, user_text, allow_fallbacks=True)
    print("route:", route, "model:", out.model, "usage:", out.usage)
    return out.choices[0].message.content or ""
 
print(agent_step("What is 2+2?"))

Application routing multiplies savings beyond provider sort alone.
Keep the classifier prompt short; it still costs tokens.
Always retain fallbacks on the expensive path so HARD tasks still complete.

Related: How OpenRouter's Provider Fallback Actually Protects Uptime

What to Practice Next

Tighten only / ignore lists once you know which providers pass evals.
Add latency sort for interactive agents.
Wire spend alerts from usage.cost and the Activity or Analytics APIs.

Related: Comparing OpenRouter's Per-Model Pricing Against Direct Provider Rates

Related: OpenRouter Cost & Fallback Best Practices

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.

Cost & Fallback Basics

Buscar en la documentación

Cost & Fallback Basics