The $20 OpenRouter Insurance Policy For Production AI Agents

TLDR: Pawel runs an autonomous agent overnight, and one morning it logged "supported providers exhausted" at 04:17. His fix is a five-rung fallback stack with $20 parked in OpenRouter, doubling as both insurance against provider outages and an extension cord for capabilities he refuses to bake into the primary architecture. The mechanism is about forty lines of Python.

Summary: I have been reading a lot about local LLMs lately and there is a quiet assumption running through that whole conversation. The assumption is that local will eventually be good enough to be the fallback when your frontier provider has a bad morning. Pawel takes the opposite position and I think he is right. Local is fine for narrow, bounded work like classification or redaction. It is not fine for the kind of multi-step reasoning you trust an Opus to do while you are asleep. Routing a sensitive task to a small local model just because the cloud is down is a "completes with a wrong answer" outcome, which is worse than failing cleanly.

The numbers in this piece are sobering if you have not done the arithmetic. Claude API's published 90-day uptime is around 98.99 percent. Sounds great. That is roughly 21 hours of downtime in a quarter. If your agent runs on a cron schedule it is going to run during some of those hours, and on the Tuesday in May 2026 when Anthropic and Gemini were both having a hard time at once, you would have wished for a second vendor in the stack. This is the bit that gets glossed over. Vendor outages tend to be full-vendor outages. Hopping Sonnet to Haiku does not help you when the Anthropic surface as a whole is unhealthy. Hopping to a different lab does.

The architecture itself is five rungs and it is the kind of thing I would have wanted to read three years ago. Primary call picks the model by task stakes. In-provider cascade catches transient failures within one vendor. Cross-provider hop swaps to a different lab through a different CLI. OpenRouter degraded mode kicks in when both vendor CLIs are unreachable, prefixing the response with "[Fallback Mode]" so the user knows what they are reading. Queue and retry catches anything left and stamps it for the next cycle. The agent never just drops a task. The honesty of that "[Fallback Mode]" prefix is the part I want to underline. The worst pattern in production AI work is a silent quality drop where the user thinks they got Opus and actually got a 9B open-weight model. Tell the truth, every time.

The cost discussion is where the piece earns its keep for me. Llama 3.3 70B on OpenRouter sits at about $0.10 input and $0.32 output per million tokens versus Sonnet 4.6 at $3 and $15. That is roughly 30x cheaper on input and 47x cheaper on output for work that does not need the frontier. Classification, redaction, the "is this email worth waking me up" pass, none of these need a top model. The myth about OpenRouter carrying a meaningful Claude markup over Anthropic-direct died with Sonnet 4.6 and Opus 4.7 pricing matching the Anthropic API exactly. If you have not checked recently, check.

What I appreciate most is the extension half of the argument. The $20 is not just there for the bad days. On the good days it is an extension cord for capabilities Pawel deliberately keeps off the primary architecture. Image generation through Nano Banana. Long-context refactors that do not fit Claude's window. Cheap evals comparing three models on the same prompt without standing up three SDKs. Each of those could be folded into the primary stack and none of them earn that complexity. Living on the extension side keeps the core agent loop small.

Key takeaways:

A single-provider agent is one bad Tuesday away from going dark. 98.99% uptime is 21 hours of downtime per quarter.
Cross-provider hops matter more than in-provider cascades because most outages are full-vendor outages, not model-specific.
Trigger conditions for fallback are timeouts, 5xx errors, persistent 429s, and 401 after one re-auth attempt. A 400 is your bug, not theirs.
Tell the user when they are in fallback mode. Silent quality drops are the worst failure mode in production agent work.
Reserve frontier models for work that earns the spend. Llama 3.3 70B handles classification at roughly an order of magnitude less cost than Sonnet.
Test by killing the primary key for ten minutes during a scheduled run. If your fallback path was never exercised, it does not work.
Cache system prompts. Tailor the identity prompt budget to the fallback model's context window.

Why do I care: Most of the agent code I read in the wild has a hard dependency on one SDK and one vendor. That dependency is going to bite, and the day it does the migration is going to be a sprint rather than an afternoon. The five-rung pattern here is portable to frontend agent work too. If you are building a Next.js app that talks to an LLM, the same logic applies. Wrap the call. Define the trigger conditions. Surface the degraded mode in the UI so users are not lied to by a silent fallback. The economics of routing classification and redaction passes to a cheap open-weight model through OpenRouter while keeping Opus for the work that actually needs it is the kind of architectural lever that compounds. I am going to take the cost gate idea and apply it to the AI features I have been bolting onto our marketing site, because right now every call goes to the most expensive model in the stack regardless of what the task actually needs.

My AI Agent's $20 Fallback Mechanism: Half Insurance, Half Extension