Most developers wiring an AI coding agent up to multiple model providers land on OpenRouter first. It is the default answer to "how do I use Claude, GPT, and Gemini through one API key," and for a lot of use cases it is a perfectly reasonable one. But OpenRouter is not the only gateway, and for people running coding agents specifically — Cursor, Claude Code, Windsurf, and similar tools that fire dozens of requests per work session — its pricing model can work against you in a way that is easy to miss until the bill arrives.
This post walks through how OpenRouter's pricing actually works, why that model interacts badly with coding-agent usage patterns, and what a flat-rate alternative built for coding tools looks like in practice.
How OpenRouter's pricing works
OpenRouter is a routing layer in front of dozens of model providers. You send one request, it forwards to whichever provider and model you specify, and it bills you based on the underlying provider's per-token rate plus its own service margin. That is a genuinely useful trade: one integration, one API key, access to nearly every model on the market, without negotiating separate accounts with Anthropic, OpenAI, and Google individually.
The catch is that "per-token, plus a margin" is still per-token. Your bill scales with how much text goes in and comes out, not with how many times you called the API or how useful the result was.
Why per-token billing hurts coding agents specifically
A chat app where a human types a question and reads one answer is a light, predictable workload. A coding agent is not that. Tools like Claude Code and Cursor routinely:
- Re-send large chunks of file context on every turn, because the model needs to see the code it is editing.
- Loop several times per user action — plan, edit, self-review, re-edit — each loop a fresh request with its own token cost.
- Run continuously in the background during a work session rather than firing once per human question.
None of that is a flaw in the tools. It is just how effective coding agents behave. But it means the same per-token rate that is cheap for occasional chat use becomes unpredictable and often expensive once an agent is looping over your codebase for hours a day. Teams frequently do not notice this until a month where an agent got stuck in a longer-than-usual loop and the invoice reflects it.
OpenRouter vs. a flat-rate, coding-tool-native gateway
| OpenRouter | APIClaw | |
|---|---|---|
| Billing model | Per-token, passthrough rate plus service margin | Flat monthly price, daily request allowance |
| Cost predictability | Varies with context size and loop length | Fixed regardless of context size or loop length |
| API format | OpenAI-compatible | OpenAI-compatible |
| Cursor / Claude Code setup | Supported via custom base URL | Same pattern, documented specifically for these tools |
| Model selection | Very broad, long tail of providers | Major providers (Claude, GPT, Gemini, Llama, and more) |
| Payment methods | Card | Card and crypto |
Neither model is universally "better." The trade-off is real: OpenRouter's per-token approach is the right fit if your usage is genuinely light or spiky, and its provider catalog is broader than any flat-rate competitor's will ever be. A flat-rate gateway like APIClaw is the better fit once an agent is a daily part of how you work and you want the monthly AI line item to stop being a variable you have to monitor.
Setting up a flat-rate gateway with Cursor or Claude Code
Switching is a configuration change, not a rewrite, because both approaches speak the OpenAI-compatible format. For APIClaw specifically, it is a base URL and an API key:
export ANTHROPIC_BASE_URL="https://apiclaw.biz/v1"
export ANTHROPIC_API_KEY="sk-your-apiclaw-key"Claude Code picks up those environment variables directly. Cursor and most other tools expose an equivalent "custom API base URL" field in their model settings. The full walkthroughs are here:
When OpenRouter is still the right call
To be fair to it: if you use AI tools lightly, need a model that only a long-tail provider offers, or want the widest possible catalog for experimentation, OpenRouter's per-token model means you only pay for what you actually use, and its provider coverage is hard to match. It is not a bad product; it is a different pricing philosophy aimed at a different usage pattern.
The bottom line
If an AI coding agent has become a daily tool rather than an occasional one, the question worth asking is not "which gateway has the most models" but "which pricing model matches how I actually use it." For steady, heavy, agentic usage, a flat-rate AI API with unlimited tokens per request removes the per-token variable entirely — you pay the same amount whether today's session was quiet or your agent looped forty times on a gnarly refactor.