Know what your AI costs.
Per model, per machine, per month.
Our agents call a lot of models across a lot of providers. Until recently we could not say what last month cost, which model drove it, or when prepaid credit was about to run out. ToggleLogic Free now records every model call in dollars, and we run our own business on it. Here are our real numbers.
These are real figures from our own headquarters agent system, September 3 to October 10, 2026. Costs are estimates at each provider's published rates, which is what ToggleLogic Free records. They are not provider invoices.
An agent bill you can't explain is a bill you can't control.
A working agent calls several models from several providers every day. Each provider bills separately, in tokens, on its own schedule. Nobody can tell you what last month cost per customer or per machine, and prepaid credit runs out without warning, usually in the middle of a job.
Every model call, in dollars, on the record.
Dollars, not tokens
Each call is recorded with the machine it ran on, the provider, the model family and exact model, and its estimated cost in dollars at the provider's published rate.
Follow a request across models
When one request fans out into several model calls, every call carries the same request ID, so you can add up what a single job really cost.
No silent $0.00
Every cost says where its price came from. A call with no known price, or no usage reported, is counted separately and flagged. It never shows up as free.
Ready for your bookkeeping
Each call says whether it is billable and why, so internal use and customer use can be separated before anything turns into a statement.
Totals, alerts and a way around empty accounts.
- A monthly export. One command writes a month of calls to CSV or JSON for billing or bookkeeping.
- Budget alerts. Set a monthly dollar limit and ToggleLogic records an alert as spend crosses 50%, 80% and 100% of it.
- Prepaid balance estimates. Record your top-ups and ToggleLogic estimates what is left from the spend it has seen, and warns once when a balance runs low. It is an estimate, and it says so.
- Out-of-credit providers are routed around. When a provider refuses calls for billing reasons, ToggleLogic records it, pauses that provider and sends later turns elsewhere. Your own overrides still win.
- Fallback by model family. You can list model families in order of preference, and ToggleLogic uses the newest model you have configured in each. It is off until you turn it on.
Our own numbers, model by model.
Motherboard runs a small fleet of AI agents: our own headquarters system plus agents at customer sites. We pay for AI usage centrally, and the ToggleLogic Free record is what we use to account for it. The figures below are our headquarters system only.
| Model family | Calls | Share of calls | Estimated cost | Share of cost |
|---|---|---|---|---|
| Google Gemini Flash | 682 | 63% | $60.30 | 65% |
| xAI Grok | 217 | 20% | $12.95 | 14% |
| OpenAI GPT | 51 | 5% | $14.05 | 15% |
| Anthropic Claude Haiku | 83 | 8% | $3.51 | 4% |
| Anthropic Claude Sonnet | 16 | 1% | $1.49 | 2% |
| Anthropic Claude Opus | 1 | <1% | $0.71 | 1% |
| Local model (on our own hardware) | 39 | 4% | $0.00 | 0% |
| Total | 1,089 | 100% | $93.01 | 100% |
104 calls have no price: the 39 local-model calls, which cost nothing in cloud fees, and 65 calls that a provider refused or that reported no usage. They are counted, not hidden. Cached context accounts for $24.86 of the total.
Same calls, same input, output and cached tokens, priced at each model's public rate. Sending everything to one premium model would have cost about three times what we spent. Our cheapest cloud model alone would have been a little cheaper still, but we send harder jobs to stronger models on purpose. Our headquarters system chooses models with ToggleLogic Intelligence on top of Free; Free is what records the cost.
Weekly spend tracks how much work the agents did. The peak week had the most calls.
On October 7 a provider answered calls without reporting usage. ToggleLogic raised one event for that model in that hour and kept counting every call in the ledger.
What we found by running it ourselves.
We would rather you hear it from us.
- Our first month-end run found calls not marked billable. On October 1 our bookkeeping run could not tell customer usage from our own. Every call now says whether it is billable and why. That fix is now part of ToggleLogic Free.
- Early records overstated cached context. Until mid-September the ledger priced cached tokens at the full input rate. For those weeks the figures above are re-priced at each provider's published cache rate. Unadjusted, the records showed about $57 more.
- A provider ran out of credit and we only saw it later. On September 28 one provider refused about 50 calls because prepaid credit was used up. The ledger showed them as calls with no usage. ToggleLogic Free now recognizes billing refusals, pauses that provider and routes later turns around it.
Our goal is 90%.
Today we spend about 67% less than a premium-only setup would. Our target is 90%. These numbers were collected on our previous headquarters machine. We have moved headquarters to a newer machine with much more memory, so we can run a larger local model and move more routine work off paid cloud models. Local calls already cost us nothing in cloud fees. We will update this page with measured results, not projections.
Free records and exports. Your system acts on it.
ToggleLogic Free gives you the record, the export and the alerts. Turning them into invoices, emails or approvals is your own system's job; for us that is our own agent system. Automatic model selection is ToggleLogic Intelligence, a separately licensed product.
ToggleLogic Free for OpenClaw
Supported OpenClaw versions are listed in the release notes. Source-available under the ToggleLogic Free License.