Forecast an OpenClaw model switch before you change the agent

Quick answer

To forecast a model switch, keep your agent's workload fixed and change only the model price. Count input, cached input, and output separately if your provider reports them. Compare the new estimate with the old one, then test whether the replacement still completes the same tasks. A lower price per token is not necessarily a lower cost per accepted task.

This is a planning exercise, not a promise about your next invoice. Clawback's free calculator lets you change the primary, heartbeat, and sub-agent models independently. Here's a way to use it without treating its defaults as measurements.

Make a baseline you can reproduce

Pick a representative period from your own provider usage records. Record the model, request count, input tokens, output tokens, any cached-input category, and successful task count for the work you actually plan to move. Keep heartbeat and sub-agent work separate from interactive messages. If you have no usage export, label the numbers as assumptions and revisit them after a week.

For a single workload, start with this worksheet. All token counts refer to the same period:

1. Ordinary input cost = ordinary input tokens / 1,000,000 × the model's input price per million.
2. Cached input cost = cached input tokens / 1,000,000 × its applicable cached-input price. If no cache category exists, don't invent a discount.
3. Output cost = output tokens / 1,000,000 × the model's output price per million.
4. Add those amounts, then divide by accepted tasks. Repeat using the candidate model's measured token counts and current rates.

Get the rates and billing rules from your provider's current price sheet. The prices shown inside a static calculator can lag a provider change. Don't use a model label in a tool as evidence that your account gets a particular rate.

Walk through the decision in Clawback

Open the calculator and enter your primary model, estimated messages per day, input tokens per message, heartbeat frequency and model, sub-agents per day and model, and channel count. Write down the monthly estimate. Change only the primary model selector and write down the second estimate. The difference is the tool's scenario delta, not a predicted bill.

Now run a second scenario for your heartbeat model and, if relevant, a third for sub-agents. Don't change all three at once. Separate runs show which assumption moves the estimate. A one-agent switch does not mean every scheduled job or child agent inherits the new model.

Know what the calculator assumes

We checked the calculator's calculation logic for this walkthrough. It adds a fixed context allowance per message, estimates output as a fraction of the entered message tokens, assigns fixed token counts per heartbeat and sub-agent spawn, multiplies by 30 days, and scales each category by the channel count. Those are calculator assumptions, not observations of your deployment. It doesn't import your provider bill, measure cache hits, or know which requests were retried.

If your workload spans several channels, compare the tool's channel multiplier with your actual traffic. A second channel does not automatically double every kind of request. If you have provider usage records, calculate the worksheet from those records before you approve a budget. Use the interactive tool to compare scenarios, not to certify a charge.

Run a small quality check before switching

Choose a handful of representative tasks and define what an accepted result means before testing. Run the same tasks on the current and candidate model. Log attempts, retries, total tokens, and accepted results. If the candidate needs more retries or fails important tasks, use its observed token mix and success rate in the worksheet instead of assuming identical volume. Keep the old model for tasks where the trade-off doesn't hold.

Want to map a different mix? Try the free calculator for a quick baseline and one-variable scenario. Paid spend tracking isn't live here; the early-access page currently opens an email request, not an account signup.