Pick a model and plan the bill with cost sweeps

Model choice and cost belong together: a cheap model is useless if it lacks tool calling, and a great model is useless if the bill doubles with traffic. The RAG assistant: grounding and cost template has both checks.

1. Model requirements

Select Tools + structured outputs, ≥128k context. The Model requirements check runs the AI Model Compatibility Explorer:

• Model data: Live list from OpenRouter or a pasted OpenRouter models JSON. The template uses the dated snapshot that ships with the tools (see the check's notes) so the numbers are reproducible; switch to live for today's list.
• Required input types, Required API features (tools, structured_outputs …), Minimum context window and Minimum max output.

The result: 5 of 6 models meet the requirements, with a table of context sizes and prices per million tokens and the Near misses that fail only one requirement.

2. From models to a cost scenario

Below the table, Price these models in a new cost scenario creates a Cost scenario with the matching models and their prices from this list. Set your traffic — Requests per day, Input tokens per request, Output tokens per request — and optionally a Monthly budget. Under Caching and batch discounts you set the cacheable prompt prefix, the cache hit rate and the batch share.

3. Read the cost result

The template's scenario RAG workload, 20,000 questions a day gives Cheapest model meeting the constraints: mistralai/mistral-small-3.2-24b-instruct at 604.2 USD per month. The bar chart shows every model; grey bars miss the context window or the budget of 800 USD (gpt-4o-mini 941, deepseek-chat-v3.1 1,557, gemini-2.5-flash 2,128 USD per month).

4. Cost sweep

Traffic estimates are guesses. Run cost sweep (6 traffic levels) runs the LLM Cost Explorer at 0.25×, 0.5×, 1×, 2×, 4× and 8× of your requests per day — six real tool calls — and draws the monthly cost of every model on a log scale, with your budget as a dashed line. The table below the chart lists the same numbers. Export the chart with Export → Cost sweep (SVG) or (PNG) while the check is selected.

5. Scenario A/B

Duplicate the cost check (Ctrl+D), change one assumption — for example a higher cache hit rate or a different price list — and run both. In Compare pick the two runs: the metric table shows the change in monthly cost and caching savings, and Monthly cost per model (USD) lists A, B and the difference per model.

6. Keep it honest

Prices change. The gate Monthly cost of the cheapest fitting model ≤ budget is only as good as the price list, so note where the prices came from (the check's Notes) and re-run with live prices before a decision. The cross-check on the dashboard warns when a cost scenario's cheapest model does not meet the model requirements.