Using OpenRouter?

OpenRouter routing or fallbacks misbehaving? We ship it properly.

OpenRouter is the easiest way to treat frontier models as commodities, and the routing choices you make there show up directly in your latency, cost and reliability curves. When a fallback chain picks the wrong provider, a provider pricing change catches you by surprise, or rate limits land somewhere you did not expect, we untangle it and set the routing to reflect what you actually want.

  • Since 2015Fixing technology
  • FreeEstimates
  • US teamReal engineers

openrouter.log

still broken

What it is

OpenRouter at a glance

OpenRouter is a unified API gateway for large language models. It exposes a single OpenAI-compatible endpoint that routes requests to Anthropic, OpenAI, Google, Mistral, Meta, DeepSeek, Groq, Together, DeepInfra and dozens of other providers, picking among them by cost, latency or availability.

You pay OpenRouter per token at roughly the underlying provider cost plus a small margin, and in return you get fallbacks, A/B routing, provider preferences and a single invoice instead of eight. The service is popular with teams that want to experiment across models cheaply, with agent frameworks that want provider abstraction, and with production applications that need failover when one vendor has a bad afternoon.

Type
LLM API gateway and routing service
First released
2023
License
Proprietary SaaS, pay-per-token
Maintained by
OpenRouter.ai
Typically used for
Multi-model routing, cost optimisation, provider failover, unified LLM billing

The honest version

OpenRouter pros and cons

No technology is all upside. This is where OpenRouter earns its place and where it tends to cause trouble.

Where OpenRouter is strong

  • OpenAI-compatible APIExisting code written against the OpenAI SDK points at OpenRouter by changing one base URL, and most calls work unchanged across dozens of providers.
  • Granular provider preferencesYou can prefer cheapest, fastest, or a named provider order, and set data policies that exclude providers who log prompts.
  • Automatic fallbacksWhen a provider is rate limited or down, OpenRouter retries on the next option in the configured list, which smooths over individual outages.
  • Unified billingOne invoice and one credit pool across every supported provider removes a lot of finance plumbing when experimenting with new models.
  • Access to models before direct integration shipsNew Meta, Mistral, Google and open-weights releases tend to appear on OpenRouter within hours, well before every SDK exposes them.
  • Streaming works across providersServer-sent events are normalised, so the client stream format stays stable even when the underlying provider changes mid-request.

Where OpenRouter bites

  • You are paying a marginOpenRouter adds a small percentage on top of raw provider cost. At low volume that is a bargain; at very high volume direct contracts with the top one or two providers may win.
  • Provider feature gaps leak throughTool calling, structured output, prompt caching and vision are not implemented uniformly across providers, so a request that works on one route fails on another.
  • Rate limits are shared across routesA fallback does not always escape the limit, because the upstream provider's quota is per account not per request, and a loop of retries can exhaust it.
  • Observability is limitedYou can see which provider served a request in the dashboard, but detailed traces and cost attribution usually need a separate layer like Langfuse.
  • Data policy depends on the routeSome providers behind OpenRouter log and train on prompts by default, which is fine for experiments and not fine for regulated data unless you filter them out explicitly.
  • Model availability driftsModel slugs appear, rename and disappear as providers come and go, so hard-coded model IDs need review periodically.

Sound familiar?

Common OpenRouter problems we fix

The OpenRouter problems that bring people to us, and what you have once each one is fixed.

429 rate_limited: provider upstream limit

Primary provider is rate limited

Traffic is hitting the Anthropic or OpenAI quota behind OpenRouter rather than a routing limit, and fallbacks are not configured to escape it cleanly. Retries pile up and the end-user call eventually times out.

Fallbacks configured, limits respected

model removed, no auto-mapping configured

A provider dropped the model

Model slugs come and go as vendors retire checkpoints. If the application pins a slug that has been removed, requests fail outright rather than quietly moving to the replacement.

Alias set, requests continue

tool_choice not supported by provider

Fallback provider lacks tool calling

The primary model supports tool use, the fallback does not, and a request that fell through on rate limits hits the second route where tool_choice is invalid.

Routes grouped by capability

402 Insufficient credits

Credits ran out mid-day

Auto top-up disabled or a low balance that drained during a traffic burst. Every request fails until the account is funded, and nobody set up a balance alert.

Billing reliability restored

provider_preferences.order ignored

Routing is not honouring the configured order

A conflicting query parameter (route=:nitro, for instance) or a preferences object that was not merged, so requests end up at the cheapest-fastest option instead of the configured provider list.

Routing behaves as configured

503 provider unavailable

Upstream outage with no failover

A single provider misconfigured as the only option for a model, so a routine outage takes down every request for that route rather than moving to a backup.

Every model has a fallback path

latency 11,400ms, routing away

Slow provider degrading tail latency

An otherwise cheap provider is running hot, and requests that land on it stretch p99 latency past user tolerance. Preferences need to weigh latency alongside cost.

Routing preferences include latency

price change: $5.00 -> $7.50 per 1M tokens

A provider raised prices quietly

The underlying cost changed, and the application is still routing to it as the cheapest option out of habit. The bill moves without anyone noticing.

Pricing watched, routes resorted

What we do

OpenRouter help, from routing to cost tuning

Pull us in for one bad fallback, or let us design the full multi-provider strategy.

Routing design
Provider preferences, fallback chains and data policies set so the primary, backup and emergency routes all match your reliability and compliance needs.
Cost tuning
Model mix reviewed against real prompts, with cheaper routes for easy traffic and reserved capacity on the premium models for the high-value calls.
Failover testing
Fault-injected runs to confirm the fallback actually kicks in cleanly when a provider returns 429 or 503, rather than failing on the first upstream error.
Observability integration
Langfuse or LangSmith wired in so you can see which provider served each request and what it cost, instead of relying on monthly aggregates.
Migration from direct APIs
Moving code off direct OpenAI, Anthropic and Mistral SDKs onto a single OpenRouter integration without losing feature parity.
Managed Tech
Routing kept current as model slugs move and prices drift, with alerts on credit balance, latency spikes and provider feature regressions.

How it works

From first message to fixed

  1. Look Free estimate

    Tell us what's going wrong. We look at it, then tell you plainly what's wrong and what it will cost.

  2. Fix

    We fix the cause, not the symptom. If something is down, getting it back up comes first.

  3. Harden

    Backups, monitoring, updates and access, so the same problem can't quietly come back.

  4. Hand back or stay on

    You get a plain-language write-up. Or we stay on for hosting, upkeep and new work.

Often found together

Running OpenRouter alongside something else?

We work on the whole stack, so the problem can be wherever it actually is.

Questions

OpenRouter questions

Something else? Call+1 737 727 7327

Why are my OpenRouter fallbacks not kicking in?

The common patterns are a fallback list that contains only one real option, a route override in the URL or query string that bypasses preferences, and shared rate limits upstream that take the whole account offline rather than just the primary provider. Check the actual preferences object sent with the request, then look at which provider the dashboard says served each failed call.

How do we keep OpenRouter costs predictable?

Set hard spend caps and credit alerts, prefer cheaper models for the majority of traffic with an escalation path to premium only when needed, and review the model mix monthly because prices move. Also tag requests so you can attribute cost to features and shut off the ones that are not worth it.

Can OpenRouter replace direct provider contracts entirely?

For most teams, yes. At very high volume, direct contracts with the one or two providers you rely on most will beat OpenRouter on price, but you lose the failover and the single-bill simplicity. A hybrid is common: direct for the hot path, OpenRouter for everything else.

Which data flows through which provider on OpenRouter?

Each model route lists the providers who can serve it and their data policies. You can filter out providers that log or train on prompts, which is the setting most regulated workloads need. Review this per route rather than account-wide, because the available providers change per model.

How do we handle a model being removed from OpenRouter?

Set aliases or a routing layer in your own code that maps retired slugs to the recommended replacement, and treat model IDs as configuration rather than hard-coded strings. OpenRouter publishes replacements for most retired models, but the mapping still needs to happen somewhere.

How much does OpenRouter help cost?

Our quote is free. Send over the symptom, whether it is a failing fallback, a spend alert, or just a routing design you want reviewed, and we price the work before touching anything.

Free estimate

OpenRouter giving you trouble?

Tell us what you're running and what's going wrong. A person reads it and comes back with an estimate.

  • No charge to find out what it will take
  • If it's urgent, say so

Replies come from a person. By sending, you authorize us to email you back. No marketing.