Search "OpenRouter alternative" and you will find listicles: ten gateways, ten logos, no numbers. This is not that. If you are evaluating gateways, the honest answer is that most differences collapse into four checks you can run in an afternoon — and the fifth thing that actually matters is your own usage log.
Check 1: Is the rate card legible?
A gateway either publishes per-model per-token prices or it doesn't. If pricing requires talking to sales, that is the pricing. Look specifically for:
- Input, output, and cached input as separate numbers — agent workloads live and die on the third.
- The official list price printed alongside — the honest measure of what the gateway's markup actually is. At AtmoRouter every model page shows both, and the delta is the whole story.
- No seat fees, minimums, or expiring credits. Deprecation of your prepaid balance is a business model, not a feature.
Check 2: How are failed requests billed?
Run any agent framework for a week and count retries: upstream hiccups, timeouts, malformed tool calls. Gateways split into two camps — those that bill attempts, and those that bill successes only. At 5-10% retry rates on chatty agent traffic, this alone is a double-digit percentage of your bill. AtmoRouter never bills a request that errors; if a gateway's policy is unstated, assume the worse option.
Check 3: Protocol coverage
Your tooling speaks OpenAI Chat Completions, Anthropic Messages, or both. A gateway that serves both protocols over its whole catalogue keeps your options open in both directions — OpenAI-flavored tools can reach any model, and so can Claude-flavored ones. This sounds like a nicety until your favorite agent framework hard-codes one dialect.
Check 4: Catalogue honesty
Two failure modes to probe:
- Ghost models — listed, priced, but effectively down. Before committing, hit each family you care about with a handful of real requests and look at error rates, not status pages.
- Inflated "discounts" against official list prices. The check is arithmetic: divide the gateway's price by the provider's published rate. A "90% off" claim that survives division is real; one that doesn't was marketing.
The fifth thing: your own usage log
Take your last month of real requests, price them at each candidate's rate card, and compare. This is the only comparison that is not a listicle. The result depends entirely on which models you actually call — a gateway that wins on the models you use is the one that wins, full stop. Free models are the cheapest way to validate a gateway end-to-end before you commit traffic: run your integration against something $0, and only then move it to the paid tier.
Where AtmoRouter lands on its own checklist
Published per-token rates next to official prices, failed requests never billed, both OpenAI- and Anthropic-compatible endpoints, and a catalogue we probe continuously — that is the design brief. Where we are not the right answer: if you need a provider so long-tail it is not in the catalogue, use the directory-style gateway; that is what they are best at.