The short answer
Compare the cost of completing a task correctly, including retrieval and maintenance as well as model usage.
Estimate model usage, indexing, storage and maintenance separately, with a transparent arithmetic example and a worksheet for your own workflow.A smaller prompt is only one part of the bill
Sending a few relevant passages can reduce a request's input, but retrieving those passages may add indexing, query processing and review. Retries cost something too. Compare the cost of a completed, correct task before deciding that the smaller prompt saves money.
Keep acquisition costs, recurring usage and human maintenance separate. A downloaded software licence and a hosted model account pay for different things. Use your actual provider invoices and observed task traces rather than assuming a tutorial's example rates match your setup.
List the cost categories before comparing plans
| Category | What to measure |
|---|---|
| Software or service | Initial licence, subscriptions and applicable updates. |
| Model usage | Input, output and any separately billed usage classes. |
| Retrieval and indexing | Initial processing, changed-source work and queries. |
| Storage and operations | Backups, hosting, monitoring and recovery. |
| Human review | Setup, corrections, source maintenance and failed-task recovery. |
Some categories may already be covered by tools you own. Others may not appear on a cloud bill because they consume your time. Record those assumptions explicitly instead of counting every existing cost as a new purchase.
Work through a deliberately simple example
Assume 100 tasks. In approach A, each task sends 12,000 input tokens and receives 800 output tokens. In approach B, each sends 3,000 input tokens and receives the same 800 output tokens. Use hypothetical rates of $1 per million input tokens and $4 per million output tokens.
| Illustrative calculation | A | B |
|---|---|---|
| Input tokens | 1,200,000 | 300,000 |
| Output tokens | 80,000 | 80,000 |
| Input cost | $1.20 | $0.30 |
| Output cost | $0.32 | $0.32 |
| Total for these two categories | $1.52 | $0.62 |
The difference is $0.90 across 100 tasks, before other costs. These are invented inputs for arithmetic, not current provider prices, measured Gigabrain usage or a promised saving. The worksheet includes a runnable calculation.
Add the costs the example leaves out
If approach B incurs $2 of additional processing in the same period, its modeled total becomes $2.62. That costs $1.10 more than A under these assumptions. The extra processing reverses the saving in this example. Input-token reduction alone cannot establish financial value.
Include retries and review time for incorrect answers. Also distinguish initial indexing from recurring work. Spreading a one-time setup cost over a month or a year changes the comparison, so state the period and expected task volume.
Measure a representative task before scaling
Choose several real task types, record the calls needed to complete them and inspect the result for correctness. Count the full workflow, including retrieval and recovery after failure. Separate cases that already fit comfortably in context from those requiring durable project knowledge.
Count time saved only when it reduces actual work. Record actual review effort and whether the result was usable. With a small pilot, report the cases and assumptions rather than projecting a precise annual return.
Compare usage alongside the software price
The cost worksheet separates usage from software and operating responsibility. Gigabrain's pricing page is the source for current package terms. Model-provider charges and your maintenance effort are separate.
Read the build-versus-buy comparison if you are deciding whether to assemble the system yourself. Weigh the price against the setup and workflows you would otherwise need to build. If Gigabrain looks like a fit, test that judgment on one small project using a supported agent.
Use the measured run as an input record
On 7 September 2026, the recorded retrieval experiment used Python 3.11.14 on macOS 14.7.6. It reused a local model and runtime seed with the Full 1.1.0 archive.
Two fictional corpora, seven queries and four configurations produced 56 retrieval runs. No remote authoring model was requested. This does not make the machine, initial model setup or operator time free.
| Input | Recorded boundary |
|---|---|
| Retrieval runs | 56, with raw per-query output. |
| Source data | Two small fictional corpora, available for inspection. |
| Remote authoring request | None requested by this runner. |
| Hardware price, electricity and staff time | Not measured. Supply your own dated inputs. |
| Package price | Use the current pricing page at purchase time. |
The report supports a workload description. It does not support a universal cost per task or a percentage saving against another product.
Add your dated invoices and measured review time to the cost worksheet. Keep the earlier hypothetical token calculation separate.
Method note: Gigabrain Labs prepared this material with AI assistance. Fictional exercises and executed component checks have separate labels.
