Tokenminning with OpenRouter

OpenRouter is a unified LLM gateway — not an IDE. Cline, Aider, OpenCode, and custom agents send requests through your OpenRouter API key. Waste shows up as frontier defaults, missing fallbacks, unstable prefixes (no cache hits), and shared keys without caps.

Work through the steps below in order. For the general technique stack, see Where to start. Pair this guide with your editor guide — OpenRouter forwards what clients send.

Quick checklist

1. OpenRouter Activity — spend by model and API key.
2. Set per-key spending limits and workspace budgets.
3. Put cheap models first in fallback arrays — routing.
4. Stabilize system prompts for prefix caching.
5. Log usage — cached_tokens confirms cache hits.
6. Enforce org guardrails before production traffic.

Typical impact: 50–90% routing routine steps to haiku-tier; 10–30% from cache-aware prefixes; caps prevent runaway agent loops.

How to reduce OpenRouter spending

1. Measure first

Open openrouter.ai/activity . Filter by API key and model. Export CSV weekly. For one expensive generation, call GET /api/v1/generation?id=... to reconcile token breakdown.

→ Activity & keys

2. Cap spend

Create separate API keys per tool (dev, CI, prod). Set a spending limit on each key at creation. For teams, use workspace budgets  and org guardrails .

→ Activity & keys · Guardrails

3. Route cheap

Set a haiku-tier primary model and a capable model in the models fallback array. Add provider.sort: "price" for cost-sensitive loops. Avoid openrouter/auto and :nitro unless you have budget for unpredictable tiers.

→ Model routing

4. Cache prefixes

Stabilize system prompts and tool schemas across turns. Read usage.prompt_tokens_details.cached_tokens after turn 3+. Set a stable session_id for sticky provider routing. Enable response caching  for identical requests.

→ Prompt caching

5. Trim upstream bloat

OpenRouter forwards what clients send — trim in each tool:

6. Guardrail in production

Use model allowlists, workspace isolation, and the Batch API  for offline eval workloads. Wire the MCP server  during development to check live model prices before you ship defaults.

→ Guardrails

How OpenRouter bills

Provider token cost + small platform fee. Prepay credits or invoice billing.

Each request: model, optional models fallbacks, optional provider sort (price / throughput / latency).

Response usage includes prompt/completion tokens, cached_tokens, cache_write_tokens, and cost. Failed fallbacks are not billed.

Guides

Connect your stack

Point any OpenAI-compatible client at https://openrouter.ai/api/v1:

import { OpenRouter } from "@openrouter/sdk";
 
const client = new OpenRouter({
  apiKey: process.env.OPENROUTER_API_KEY,
});
 
const completion = await client.chat.send({
  model: "anthropic/claude-haiku-4-20250514",
  messages: [{ role: "user", content: "Summarize this diff." }],
});

See the OpenRouter quickstart  for Python, curl, and Agent SDK patterns.

Troubleshooting

When gateway tuning is not enough, instrument production agents with per-feature tags and Context hygiene runtime caps.

Anti-patterns

Related

• Gateways
• Prompt caching
• OpenRouter docs