Start by measuring what you pay for each API request that works. Use a mix of short questions, long chats, and requests that use tools.
Work out what a successful request costs
Record the model ID, uncached input, cached input, output tokens, and final charge. Add up all charges, including paid failures and retries. Divide by the number of successful requests. That’s your starting point.
Put instructions and tool definitions that stay the same at the start of the prompt. This shared opening, called a prefix, gives the provider a chance to reuse work. Run the same tasks again and check the bill. Compare cache reads and track cache writes separately if your provider reports them.
For example, the Jarbird pricing feed quotes $0.0753 for a GPT-6.1 Sol request with 48,000 input tokens, 2,000 output tokens, and no cache hit. The equivalent OpenRouter list price is $0.1160 before its credit-purchase fee. Those are price estimates from the feed. Use your own billed requests to measure what you actually save.

Try smaller requests and compare routes
Remove old chat messages the current question doesn’t need. If your app needs short answers, try a lower output limit. Check for worse answers, cutoffs, and incomplete tool calls.
When you compare routes, use the same model, prompts, settings, and output limits. Run separate tests with uncached and cached input.
- Track success rate, time to the first token, and total cost.
- Check answer quality on the same tasks.
- Pick the cheapest route that meets your speed and quality needs.
Check the result on a sample of live traffic. Save your starting numbers so you can run the comparison again when usage or prices change.