What's a good OpenRouter alternative for production agents?
Standard Compute: its router sends each step to the cheapest model that can do it, on providers we vet for price, uptime and speed with automatic fallback, and each key's monthly budget stops new requests instead of running over. Opt-in Europe or US Compliance rules cover a key, team or company, and we don't train on your prompts or store them.
# One key per service, so spend and logs stay separate
import os
from openai import OpenAI, APIStatusError
client = OpenAI(
base_url="https://api.stdcmpt.com/v1",
api_key=os.environ["STANDARD_COMPUTE_API_KEY"],
max_retries=3, # backs off on 429 and 5xx
)
try:
reply = client.chat.completions.create(
model="standardcompute", # or an exact model ID
messages=[{"role": "user", "content": "Summarize the failing test"}],
)
except APIStatusError as err:
if err.status_code == 402: # budget used up
pause_and_alert() # the SDK won't retry it
raise
Use model standardcompute to let the router choose per request, or an exact catalog ID. Claude Code needs every model slot set to a catalog ID.
What a production agent needs from its gateway
| Need | On Standard Compute |
|---|---|
| Lower cost per task | Model standardcompute sends routine steps to efficient models and harder ones to stronger models. Pin an exact ID and the request stays on that model. |
| Providers that stay up | We set a vetted provider order for every model. When one provider fails, the request moves to the next provider of the same model. |
| Spend that stops | At a key's monthly budget, new requests get HTTP 402 budget_exhausted. Running requests finish, and other keys keep working. |
| Retries | Transient upstream 402, 429 and 5xx errors are retried up to twice, honoring Retry-After. Streams retry only before the first byte. |
| Rate limits | 300 to 4,000 requests a minute per key, by plan. Above that, HTTP 429 with Retry-After. |
| Usage per key | A request log per key with status, the model that served it, latency and finish reason. Prompt and response text are never logged. |
| Agent endpoints | Chat Completions, Responses for Codex CLI and Anthropic Messages for Claude Code, all streaming, on one key. |
| Keys for a team | A plan per key, admins and members, teams, Admin-managed keys for shared services and an audit log. |
Side by side with OpenRouter, Requesty, Portkey and LiteLLM
| Gateway | How you pay | Who picks the model | Spend limit per key | Fallback | SSO, SLA, audits |
|---|---|---|---|---|---|
| Standard Compute | Plan per key, $19 to $249 a month with $20 to $269 of compute; Business from $499 | The router per request, or you pin a model | Plan budget, then HTTP 402 | Next provider of the same model | Audit log; no SSO or SOC 2; SLA in custom Business contracts |
| OpenRouter | Provider price plus 5.5% on credits, 8% on Business | You, or its Auto and Pareto routers | Credit limit per key with a reset period; guardrail budgets | Next provider by default; a models list for model fallback | SSO/SAML and contractual SLAs on Enterprise |
| Requesty | Provider price plus 5% on top-ups; your own keys | Fallback, load-balancing and latency policies you write | Monthly cap per key; group budgets on Enterprise | Fallback chains and managed failover | SSO, audit logs and SLAs on Enterprise; SOC 2 Type II in progress |
| Portkey | Free to $49 a month for logs, plus each provider's bill | Conditional routing and weights you configure | Key and workspace budgets on Enterprise and select Pro | Fallbacks and load balancing in a config | SOC 2, ISO 27001, HIPAA; SSO, SCIM and SLAs on Enterprise |
| LiteLLM, self-hosted | Free software; you pay providers and run it | Its Auto Router add-on, by difficulty tier, on the proxy you run | Virtual-key budgets and RPM limits, with its database | Fallbacks you configure | SOC 2 Type II; SSO free for up to 5 users; audit logs and SLAs on its Enterprise license |
From each vendor's own pages, all checked October 14: OpenRouter pricing and provider routing, Requesty, Portkey and LiteLLM. Standard Compute is our product.
When OpenRouter is the better pick
- You need a model or upstream provider outside our 35-model catalog. OpenRouter lists 500+ models from 80+ providers.
- You want to bring your own provider keys. We don't support BYOK.
- Procurement needs SSO/SAML, SOC 2 or a contractual SLA today. We have none of these on self-serve plans; the status page publishes a 99.9% monthly uptime target, and an SLA comes with a custom Business contract.
- Your traffic is light or irregular, so paying per token costs less than a plan. Unused budget does not roll over.
- The whole request path must stay in the EU. See the Europe comparison.
- You want a second network path. We reach providers through OpenRouter, so a broad OpenRouter outage reaches us too.
Where Standard Compute fits
A plan per key from $19 a month, the router spending it on efficient models, and a 402 handler in your service where you already handle 429.
FAQ
Is OpenRouter reliable enough for production?
Many teams run it in production: it moves to the next provider by default and passes provider rate limits through. What usually sends teams looking is cost control, since per-token bills grow with every retry and resent context. Standard Compute uses OpenRouter upstream too; what it adds is per-step routing, a vetted provider order and a budget per key.
What happens to my agent when a key runs out of budget?
New requests on that key get HTTP 402 with the code budget_exhausted until the plan renews. The OpenAI SDKs don't retry a 402, so catch it, pause the queue and alert someone. An admin can move the key to a bigger plan, or turn on Pay as you go at usage plus 15%.
What are the rate limits per API key?
300 requests a minute on Starter, 600 on Economy, 1,000 on Standard, 2,000 on Pro and Pro Plus, 3,000 on Growth and 4,000 on Scale. Above that the key gets HTTP 429 with Retry-After and the code rate_limit_exceeded.
Can I see which service or developer spent what?
Yes, with a key per service or developer. Each key has its own usage view and request log. Admins see every company key, and the audit log records who created, rotated or revoked keys and who changed plans and rules.
How do I move a production service off OpenRouter?
Change the base URL to https://api.stdcmpt.com/v1, swap the key and set the model to standardcompute or an exact catalog ID. Streaming, tool calls and the OpenAI SDKs work as before. Move one service first and compare its cost and errors in the request log.